Proofs that require fundamentally new ways of thinking I do not know exactly how to characterize the class of proofs that interests me, so let me give some examples and say why I would be interested in more. Perhaps what the examples have in common is that a powerful and unexpected technique is introduced that comes to seem very natural once you are used to it.
Example 1. Euler's proof that there are infinitely many primes.
If you haven't seen anything like it before, the idea that you could use analysis to prove that there are infinitely many primes is completely unexpected. Once you've seen how it works, that's a different matter, and you are ready to contemplate trying to do all sorts of other things by developing the method.
Example 2. The use of complex analysis to establish the prime number theorem.
Even when you've seen Euler's argument, it still takes a leap to look at the complex numbers. (I'm not saying it can't be made to seem natural: with the help of Fourier analysis it can. Nevertheless, it is a good example of the introduction of a whole new way of thinking about certain questions.)
Example 3. Variational methods.
You can pick your favourite problem here: one good one is determining the shape of a heavy chain in equilibrium.  
Example 4. Erdős's lower bound for Ramsey numbers.
One of the very first results (Shannon's bound for the size of a separated subset of the discrete cube being another very early one) in probabilistic combinatorics.
Example 5. Roth's proof that a dense set of integers contains an arithmetic progression of length 3.
Historically this was by no means the first use of Fourier analysis in number theory. But it was the first application of Fourier analysis to number theory that I personally properly understood, and that completely changed my outlook on mathematics. So I count it as an example (because there exists a plausible fictional history of mathematics where it was the first use of Fourier analysis in number theory).
Example 6. Use of homotopy/homology to prove fixed-point theorems.
Once again, if you mount a direct attack on, say, the Brouwer fixed point theorem, you probably won't invent homology or homotopy (though you might do if you then spent a long time reflecting on your proof).

The reason these proofs interest me is that they are the kinds of arguments where it is tempting to say that human intelligence was necessary for them to have been discovered. It would probably be possible in principle, if technically difficult, to teach a computer how to apply standard techniques, the familiar argument goes, but it takes a human to invent those techniques in the first place.
Now I don't buy that argument. I think that it is possible in principle, though technically difficult, for a computer to come up with radically new techniques. Indeed, I think I can give reasonably good Just So Stories for some of the examples above. So I'm looking for more examples. The best examples would be ones where a technique just seems to spring from nowhere -- ones where you're tempted to say, "A computer could never have come up with that."
Edit: I agree with the first two comments below, and was slightly worried about that when I posted the question. Let me have a go at it though. The difficulty with, say, proving Fermat's last theorem was of course partly that a new insight was needed. But that wasn't the only difficulty at all. Indeed, in that case a succession of new insights was needed, and not just that but a knowledge of all the different already existing ingredients that had to be put together. So I suppose what I'm after is problems where essentially the only difficulty is the need for the clever and unexpected idea. I.e., I'm looking for problems that are very good challenge problems for working out how a computer might do mathematics. In particular, I want the main difficulty to be fundamental (coming up with a new idea) and not technical (having to know a lot, having to do difficult but not radically new calculations, etc.). Also, it's not quite fair to say that the solution of an arbitrary hard problem fits the bill. For example, my impression (which could be wrong, but that doesn't affect the general point I'm making) is that the recent breakthrough by Nets Katz and Larry Guth in which they solved the Erdős distinct distances problem was a very clever realization that techniques that were already out there could be combined to solve the problem. One could imagine a computer finding the proof by being patient enough to look at lots of different combinations of techniques until it found one that worked. Now their realization itself was amazing and probably opens up new possibilities, but there is a sense in which their breakthrough was not a good example of what I am asking for.
While I'm at it, here's another attempt to make the question more precise. Many many new proofs are variants of old proofs. These variants are often hard to come by, but at least one starts out with the feeling that there is something out there that's worth searching for. So that doesn't really constitute an entirely new way of thinking. (An example close to my heart: the Polymath proof of the density Hales-Jewett theorem was a bit like that. It was a new and surprising argument, but one could see exactly how it was found since it was modelled on a proof of a related theorem. So that is a counterexample to Kevin's assertion that any solution of a hard problem fits the bill.) I am looking for proofs that seem to come out of nowhere and seem not to be modelled on anything.
Further edit. I'm not so keen on random massive breakthroughs. So perhaps I should narrow it down further -- to proofs that are easy to understand and remember once seen, but seemingly hard to come up with in the first place.
 A: Although this has already been said elsewhere on MathOverflow, I think it's worth repeating that Gromov is someone who has arguably introduced more radical thoughts into mathematics than anyone else. Examples involving groups with polynomial growth and holomorphic curves have already been cited in other answers to this question. I have two other obvious ones but there are many more.
I don't remember where I first learned about convergence of Riemannian manifolds, but I had to laugh because there's no way I would have ever conceived of a notion. To be fair, all of the groundwork for this was laid out in Cheeger's thesis, but it was Gromov who reformulated everything as a convergence theorem and recognized its power.
Another time Gromov made me laugh was when I was reading what little I could understand of his book Partial Differential Relations. This book is probably full of radical ideas that I don't understand. The one I did was his approach to solving the linearized isometric embedding equation. His radical, absurd, but elementary idea was that if the system is sufficiently underdetermined, then the linear partial differential operator could be inverted by another linear partial differential operator. Both the statement and proof are for me the funniest in mathematics. Most of us view solving PDE's as something that requires hard work, involving analysis and estimates, and Gromov manages to do it using only elementary linear algebra. This then allows him to establish the existence of isometric embedding of Riemannian manifolds in a wide variety of settings.
A: My favorite example from algebraic topology is Rene Thom's work on cobordism theory. The problem of classifying manifolds up to cobordism looks totally intractable at first glance. In low dimensions ($0,1,2$), it is easy, because manifolds of these dimensions are completely known. With hard manual labor, one can maybe treat dimensions 3 and 4. But in higher dimensions, there is no chance to proceed by geometric methods.
Thom came up with a geometric construction (generalizing earlier work by Pontrjagin), which is at the same time easy to understand and ingenious. Embed the manifold into a sphere, collapse everything outside a tubular neighborhood to a point and use the Gauss map of the normal bundle... What this construction does is to translate the geometric problem into a homotopy problem, which looks totally unrelated at first sight.
The homotopy problem is still difficult, but thanks to work by Serre, Cartan, Steenrod, Borel, Eilenberg and others, Thom had enough heavy guns at hand to get fairly complete results.
Thom's work led to an explosion of differential topology, leading to Hirzebruch's signature theorem, the Hirzebruch-Riemann-Roch theorem, Atiyah-Singer, Milnor-Kervaire classification of exotic spheres.....until Madsen-Weiss' work on mapping class groups.
A: The method of forcing certainly fits here. Before, set theorists expected that independence results would be obtained by building non-standard, ill-founded models, and model theoretic methods would be key to achieve this. Cohen's method begins with a transitive model and builds another transitive one, and the construction is very different from all the techniques being tried before. 
This was completely unexpected. Of course, in hindsight, we see that there are similar approaches in recursion theory and elsewhere happening before or at the same time. 
But it was the fact that nobody could imagine you would be able to obtain transitive models that mostly had us stuck.
A: How about Bolzano's 1817 proof of the intermediate value theorem?
In English here:
Russ, S. B. "A Translation of Bolzano's Paper on the Intermediate Value Theorem." Hist. Math. 7, 156-185, 1980. 
Or in the original here:
Bernard Bolzano (1817). Purely analytic proof of the theorem that between any two values which give results of opposite sign, there lies at least one real root of the equation. In Abhandlungen der königlichen böhmischen Gesellschaft der Wissenschaften Vol. V, pp.225-48.
Not fully rigorous, according to today's standards, but perhaps his method of proof could be considered a breakthrough nonetheless.
A: I am always impressed by proofs that reach outside the obvious tool-kit. 
For example, the proof that the dimensions of the irreducible representations of a finite group divide the order of the group relies on the fact that the character values are algebraic integers. 
In particular, given a finite group $|G|$ and an irreducible character $\chi$ of dimension $n,$ 
$$\frac{1}{n} \sum_{s \in G} \chi(s^{-1})\chi(s) = \frac{|G|}{n}.$$
However, since $\frac{|G|}{n}$ is an algebraic integer (it is the image of an algebra homomorphism) lying in $\mathbb{Q},$ it in fact lies in $\mathbb{Z}.$
A: I don't know who deserves credit for this, but I was stunned by the concept of view complicated objects like functions simply as points in a vector space. With that view one solves and analyzes PDEs or integral equations in Lebesgue or Sobolev spaces. 
A: Use of Lagrange theorem (group theory) to prove Fermat's small theorem?
Use of fixed point methods (and completeness) to prove existence of solutions to differential equations?
Use of Fields theory to prove the impossibility of the trisection of the angle?
Use of group theory to prove insolvability of 5th degree equation?
A: Novikov's proof of the topological invariance of rational Pontryangin classes, for which he was awarded the 1970 Fields Medal. Fundamentally new (complicating a fundamental group to simplify geometry), and also fundamentally important. Here is what Sir Michael Atiyah had to say (as cited in the introduction to Raniski's Higher Dimensional Knot Theory):

Undoubtedly the most important single result of Novikov, and one which combines in a remarkable degree both algebraic and geometric methods, is his famous proof of the topological invariance of (rational) Pontryagin classes of a differentiable manifold...
As is well-known many topological problems are very much easier if one is dealing with simply-connected spaces. Topologists are very happy when they can get rid of the fundamental group and its algebraic complications. Not so Novikov! Although the theorem above involves only simply-connected spaces, a key step in its proof consists in preversely introducing a fundamental group, rather in the way that (on a much more elementary level) puncturing the plane makes it non-simply-connected. This bold move has the effect of simplifying the geometry at the expense of complicating the algebra, but the complication is just manageable and the trick works beautifully. It is a real master stroke and completely unprecedented.

A: Some more proofs that startled me (in a random order):
Liouville theorem to prove that Weierstrass P-function satifies the differential equation you know.
Complex methods to establish the addition law on an elliptic curve.
Cauchy's formula (for P'/P) to prove that C is algebraically closed.
Pigeon hole principle to prove existence of solutions to Fermat-Pell's equation
Kronecker's solution to the same equation, using L-functions.
Minkowski's lemma (a convex compact, symmetric, of volume 2^n contains a non trivial
integer point) and its use to prove Dirichlet's theorem on the structure of units in number fields.
Fourier transform to prove (versions of) the central limit theorem.
Multiplicativity of Ramanujan's tau function via Hecke operators.
Poisson formula and its use (for example, for the functional equation of Riemann's zeta function, or for computing the volume of SL_n(R)/SL_n(Z), or values of zeta at even positive integers).
A: "unexpected technique"
Sometimes the result itself is unexpected.  Cantor's diagonal proof (and other counterexamples), Godel's incompleteness, Banach-Tarski and nonmeasurable sets, independence results generally.
I think you want cases where the result is anticipated, but the technique seems unrelated?
A: Proving that subgroups of free groups are free requires the knowledge of topology, a completely different field which a priori does not have anything to do with groups.
A: How about Rabinowitsch's proof of the Nullstellensatz?
A: There are two ways to prove the compactness theorem for propositional logic - either using the completeness theorem and going from semantic entailment to syntactic proof, or by a topological argument in Stone spaces. The latter, I feel, is an unexpected way of doing it - but I don't know the history of the subject so I'm probably not qualified to comment whether it was fundamentally new or not. Certainly in light of Stone's representation theorem, it seems unsurprising that there could be a topological proof of a theorem in logic, and as I understand it this connection is further investigated in topos theory?
A: The Lebesgue integral seems to have been a fundamentally new way of thinking about the integral.  It's hard to prove the convergence theorems if you have the Riemann integral in mind.  I suppose there are probably many instances where you can give a computer a very ineffective definition of something and ask that it prove theorems.  Ask it to prove anything about the primes where you start with the converse of Wilson's theorem as the definition of a prime. Can the computer figure out that its definition is terrible?  Can it figure out what a prime really "is"?
A: Turing's solution of Hilbert's Entscheidungsproblem.  The new idea was to invent the Turing machine and "virtualization" (the universal Turing machine).
A: How about Serre's proof about localization of regular rings?
Let us say that a noetherian local ring $(A,m)$ is called regular if the minimal number of generators of the maximal ideal $m$ is equal to the Krull dimension of $A$.
Then, an important basic question is: Is the localization of a regular local ring again a regular local ring?
Serre's insight was that one can characterize regular local rings using homological algebra: they are exactly the rings for which every $A$-module has a projective resolution of finite length. Once you know this, the above localization problem becomes trivial. I think this was the first application of homological algebra to commutative algebra.
A: What about Euler's solution to the Konigsberg bridge problem?  It's certainly not difficult, but I think (not that I really know anything about the history) it was quite novel at the time.
A: Would the "quantum method" fit the bill here ?
"Quantum Proofs for Classical Theorems"
Andrew Drucker, Ronald de Wolf
"Erdös and the Quantum Method"
Richard Lipton
A: Lovasz's proof of cancellation in certain classes of finite structures still bewilders me;  I can only imagine that he found the proof first and then came up with the theorem afterwards.  The basic idea is to look at homomorphisms between a given structure and a sequence of other structures.  A comparison of two such sequences involving structures of
the form AxC and BxC can be taken to a comparison between A and B.  The condition that there exists a one-element substructure is used to show a certain nontriviality of the
comparison, and a few more details result in showing A is isomorphic to B if(f) AxC is isomorphic to BxC.
I should have asked Lovasz how he came up with the proof;  I am confident that most people would not be able to come close to the method independently if they were only given the theorem statement.  (Not to mention the analogous statement of unique nth roots in the same class.)
Gerhard "Ask Me About System Design" Paseman, 2010.12.09
A: Hochster and Huneke's tight closure theory to prove various theorems in Commutative algebra (Cohen-Macaulayness of rings of invariants, existence of big Cohen Macaulay algebras)?
A: The Ax-Kochen theorem about zeros of forms over the $p$-adics which was proved using model theory.
A: I think the concepts of Archimedes which are at the birth of infinitesimal calculation, as the definition of length of a circle (hence the concept of $\pi$), and how to calculate the area of ​​a circle from $\pi$. 
A: *

*Minkowski's geometric methods in algebraic number theory.

*Kolmogorov's application of Shannon's entropy to classify Bernoulli dynamic systems.

*Applications of low-dimensional geometric topology (the fundamental group) to the abstract group theory.

*Archimedes' applications of mechanics to geometry (he considered them to be but heuristics, and provided additional "purely geometric" proofs for his results, but it was, as we know now, profound mathematics, something like affine geometry, with linear coefficients of points, adding to $1$, more or less being normalized weights).

*Banach's method of applying Baire Theorem (Baire Property).


(I have one more recent too, which took specialists by surprize, but 5 is a nice number).
A: Not sure whether to credit Abel or Galois with the "fundamental new way of thinking" here, but the proof that certain polynomial equations are not solvable in radicals required quite the reformulation of thinking. (I'm leaning towards crediting Galois with the brain rewiring reward.)
P.S. Is it really the case that no one else posted this, or is my "find" bar not working properly?
A: Technically, the following are not proofs, or even theorems, but I think they count as insights that have the quality that it's hard to imagine computers coming up with them.  First, there's:

Mathematics can be formalized.

Along the same lines, there's:

Computability can be formalized.

If you insist on examples of proofs then maybe I'd be forced to cite the proof of Goedel's incompleteness theorem or of the undecidability of the halting problem, but to me the most difficult step in these achievements was the initial daring idea that one could even formulate a mathematically satisfactory definition of something as amorphous as "mathematics" or "computability."  For example, one might argue that the key step in Turing's proof was diagonalization, but in fact diagonalization was a major reason that Goedel thought one couldn't come up with an "absolute" definition of computability.
Nowadays we are so used to thinking of mathematics as something that can be put on a uniform axiomatic foundation, and of computers as a part of the landscape, that we can forget how radical these insights were.  In fact, I might argue that your entire question presupposes them.  Would computers have come up with these insights if humans had not imagined that computers were possible and built them in the first place?  Less facetiously, the idea that mathematics is a formally defined space in which a machine can search systematically clearly presupposes that mathematics can be formalized.
More generally, I'm wondering if you should expand your question to include concepts (or definitions) and not just proofs?
Edit. Just in case it wasn't clear, I believe that the above insights have fundamentally changed mathematicians' conception of what mathematics is, and as such I would argue that they are stronger examples of what you asked for than any specific proof of a specific theorem can be.
A: The use of spectral sequences to prove theorems about homotopy groups. For instance, until Serre's mod C theory, nobody knew that the homotopy groups of spheres were even finitely generated.
A: Here are two more candidates for new ways of thinking in proofs but I am not sure about the historical picture. One is Brunn sieve which led to new results in number theory. The other is Kummer's method that have led to proofs of many cases of FLT. (Frey's new way of thinking regarding FLT was already mentioned in a Roy Smith's comment.) 
A: I work in automated theorem proving. I certainly agree, in principle, that there are no proofs that are inherently beyond the ability of a computer to solve, but I also think that there are fundamental methodological problems in addressing the problem as posed. 
The problem is to come up with a solution that would not be regarded as 'cheating', i.e., somehow building the solution into the automated prover to start with. New proof methods can be captured by what are called 'tactics', i.e., programs that guide a prover through a proof. Clearly, it would not be satisfactory to analyse the original proof, extract a tactic from it (even a generic one) that captures the novel proof structure and then demonstrate that the enhanced prover could now 'discover' the novel proof. Rather, we want the prover to invent the new tactic itself, perhaps by some analysis of the conjecture to be proved, and then apply it. So we need an automated prover that learns. But anticipating what kind of tactic we want to be learnt may well influence the design of the learning mechanism. We've now kicked the 'cheating' problem up a level. 
Methodologically, what we want is a large class of problems of this form. Some we can use for development of the learning mechanism, and some we can use to test it. Success on a previously unseen test set would demonstrate the generality of the learning mechanism and, hence, the absence of cheating. Unfortunately, these challenges are usually posed as 'can a prover prove this theorem' rather than 'can it solve this wide range of theorems, each  requiring a different form of novelty. Clearly, this latter form of the question is hugely challenging and we're unlikely to see if solved in the foreseeable future. 
A: I have two  favorite examples.
A. H. Weyl's 1916 proof of the  equidistribution  (in $[0,1]$) of the sequence  $x_n=n\alpha \bmod \mathbb{Z}$, $\alpha$ irrational.   He  formulates an  even more complicated question, namely to prove that
$$ \lim_{n\to\infty} \frac{1}{n}\sum_{k=1}^n f(x_k)=\int_0^1 f(x) dx,  $$
for any  Riemann integrable $f$.  (The uniform  distribution follows by setting  $f=$ the characteristic function of an interval.)  To prove the more complicated question  he observes that the space $X$  of $f$'s satisfying the above equality is   a vector space and it is closed with respect to a natural topology.     He  then observes that    $X$  contains all the trigonometric polynomials (trivial computation) and thus $X$ must contain all the functions that can be approximated  by trig polynomials.    This implies that $X$ contains all the  Riemann integrable functions.      This a soft touch  a proof, with no brute force computation,  using ideas of functional analysis at a time when  the ideas of functional analysis  were not part  of   the mathematical arsenal.
B. Forty years later A. Grothendieck  gave  a    beautiful proof to the (then)  recently discovered   Riemann-Roch-Hizebruch formula.    He formulated  a more complicated  problem, observed that  the more complicated   problem has a rich structure    encoded in the object he invented and  now  called the $K$-theory of coherent sheaves, and then  used  functoriality to show that to prove the most general case it suffices to prove it  for   two special  classes of examples. 
A: Having read his name among the commenters, I thought I ought to mention his theorem: Monsky's theorem is proved using valuations from algebra.
A: Another example from logic is Gentzen's consistency proof for Peano
arithmetic by transfinite induction up to $\varepsilon_0$, which I 
think was completely unexpected, and unprecedented.
A: It seems that certain problems seem to induce this sort of new thinking (cf. my article "What is good mathematics?").  You mentioned the Fourier-analytic proof of Roth's theorem; but in fact many of the proofs of Roth's theorem (or Szemerédi's theorem) seem to qualify, starting with Furstenberg's amazing realisation that this problem in combinatorial number theory was equivalent to one in ergodic theory, and that the structural theory of the latter could then be used to attack the former.  Or the Ruzsa–Szemerédi observation (made somewhat implicitly at the time) that Roth's theorem follows from a result in graph theory (the triangle removal lemma) which, in some ways, was "easier" to prove than the result that it implied despite (or perhaps, because of) the fact that it "forgot" most of the structure of the problem.  And in this regard, I can't resist mentioning Ben Green's brilliant observation (inspired, I believe, by some earlier work of Ramare and Ruzsa) that for the purposes of finding arithmetic progressions, that the primes should not be studied directly, but instead should be viewed primarily [pun not intended] as a generic dense subset of a larger set of almost primes, for which much more is known, thanks to sieve theory….
Another problem that seems to generate radically new thinking every few years is the Kakeya problem.  Originally a problem in geometric measure theory, the work of Bourgain and Wolff in the early 90s showed that the combinatorial incidence geometry viewpoint could lead to substantial progress.  When this stalled, Bourgain (inspired by your own work) introduced the additive combinatorics viewpoint, re-interpreting line segments as arithmetic progressions.  Meanwhile, Wolff created the finite field model of the Kakeya problem, which among other things lead to the sum-product theorem and many further developments that would not have been possible without this viewpoint.  In particular, this finite field version enabled Dvir to introduce the polynomial method which had been applied to some other combinatorial problems, but whose application to the finite field Kakeya problem was hugely shocking.  (Actually, Dvir's argument is a great example of "new thinking" being the key stumbling block.  Five years earlier, Gerd Mockenhaupt and I, in "Restriction and Kakeya phenomena for finite fields", managed to stumble upon half of Dvir's argument, showing that a Kakeya set in a finite field could not be contained in a low-degree algebraic variety.  If we had known enough about the polynomial method to make the realisation that the exact same argument also showed that a Kakeya set could not have been contained in a high-degree algebraic variety either, we would have come extremely close to recovering Dvir's result; but our thinking was not primed in this direction.)  Meanwhile, Carbery, Bennet, and I discovered that heat flow methods, of all things, could be applied to solve a variant of the Euclidean Kakeya problem (though this method did appear in literature on other analytic problems, and we viewed it as the continuous version of the discrete induction-on-scales strategy of Bourgain and Wolff.)  Most recent is the work of Guth, who broke through the conventional wisdom that Dvir's polynomial argument was not generalisable to the Euclidean case by making the crucial observation that algebraic topology (such as the ham sandwich theorem) served as the continuous generalisation of the discrete polynomial method, leading among other things to the recent result of Guth and Katz you mentioned earlier.
EDIT: Another example is the recent establishment of universality for eigenvalue spacings for Wigner matrices.  Prior to this work, most of the rigorous literature on eigenvalue spacings relied crucially on explicit formulae for the joint eigenvalue distribution, which were only tractable in the case of highly invariant ensembles such as GUE, although there was a key paper of Johansson extending this analysis to a significantly wider class of ensembles, namely the sum of GUE with an arbitrary independent random (or deterministic) matrix.  To make progress, one had to go beyond the explicit formula paradigm and find some way to compare the distribution of a general ensemble with that of a special ensemble such as GUE.  We now have two basic ways to do this, the local relaxation flow method of Erdős, Schlein, Yau, and the four moment theorem method of Van Vu and myself, both based on deforming a general ensemble into a special ensemble and controlling the effect on the spectral statistics via this deformation (though the two deformations we use are very different, and in fact complement each other nicely).  Again, both arguments have precedents in earlier literature (for instance, our argument was heavily inspired by Lindeberg's classic proof of the central limit theorem) but as far as I know it had not been thought to apply them to the universality problem before.
A: Use of the  Hardy–Littlewood circle method towards Waring's problem. 
A: I'm surprised that nobody has mentioned the ancient greek proof of the irrationality of $\sqrt{2}$  which certainly amazed the contemporaries!
A: I'm a little surprised no one has cited Thurston's impact on low-dimensional topology and geometry. I'm far from an expert, so I'm reluctant to say much about this. But I have the impression that Thurston revolutionized the whole enterprise by taking known results and expressing them from a completely new perspective that led naturally both new theorems and a lot of new conjectures. Perhaps Thurston himself or someone else could say something, preferably in a separate answer so I can delete mine.
A: I think that Eichler and Shimura's proof of the Ramanujan--Petersson conjecture for weight two modular forms provides an example.  Recall that this conjecture is a purely analytic statement: namely that if $f$ is a weight two cuspform on some congruence subgroup of $SL_2(\mathbb Z)$, which is an eigenform for the Hecke operator $T_p$ ($p$ a prime not dividing the level of the congruence subgroup in question) with eigenvalue $\lambda_p$, then $| \lambda_p | \leq 2 p^{1/2}.$  Unfortunately, no purely analytic proof of this result is known.  (Indeed, if one shifts one's focus from holomorphic modular forms to Maass forms, then the corresponding conjecture remains open.)
What Eichler and Shimura realized is that, somewhat miraculously, $\lambda_p$ admits an alternative characterization in terms of counting solutions to certain congruences modulo $p$,
and that estimates there due to Hasse and Weil (generalizing earlier estimates of Gauss and
others) can be applied to show the desired inequality.
This argument was pushed much further by Deligne, who handled the general case of weight $k$ modular forms (for which the analogous inequality is $| \lambda_p | \leq 2 p^{(k-1)/2}$),
using etale cohomology of varieties in characteristic $p$ (which is something of a subtle and more technically refined analogue of the notion of a congruence mod $p$).  (Ramanujan's original conjecture was for the unique cuspform of weight 12 and level 1.)
The idea that there are relationships (some known, others conjectural) between automorphic forms and algebraic geometry over finite fields and number fields has now become part of the received wisdom of algebraic number theorists, and lies at the heart of the Langlands program.  (And, of course, at the heart of the proof of FLT.)   Thus the striking idea of Eichler and Shimura has now become a basic tenet of a whole field of mathematics.
Note: Tim in his question, and in some comments, has said that he wants "first non-trivial instances" rather than difficult arguments that involve a whole range of ideas and techniques.  In his comment to Terry Tao's answer regarding Perelman, he notes that long, difficult proofs might well include within them instances of such examples.  Thus I am offering this example as perhaps a "first non-trivial instance" of the kind of insights that are involved in proving results like Sato--Tate, FLT, and so on.  
A: Gromov's use of J-holomorphic curves in symplectic topology (he reinterpreted holomorphic functions in the sense of Vekua) as well as the invention of Floer homology (in order to deal with the Arnol'd conjecture).
A: Donaldson's idea of using global analysis to get more insight about the topology of manifolds.
Nowadays it is clear to us that (non-linear) moduli spaces give something new, and more than linear (abelian) Hodge theory, for example, but I think at that time this was really new.
A: Topological methods in combinatorics (started by Lovasz' proof of the Kneser conjecture, I guess). 
A: I don't know how good an example this is.  The Lefschetz fixed point theorem tells you that you can count (appropriately weighted) fixed points of a continuous function $f : X \to X$ from a compact triangulable space to itself by looking at the traces of the induced action of $f$ on cohomology.  This is a powerful tool (for example it more-or-less has the Poincare-Hopf theorem as a special case).
Weil noticed that the number of points of a variety $V$ over $\mathbb{F}_{q^n}$ is the number of fixed points of the $n^{th}$ power of the Frobenius map $f$ acting on the points of $V$ over $\overline{\mathbb{F}_q}$ and, consequently, that it might be possible to describe the local zeta function of $V$ if one could write down the induced action of $f$ on some cohomology theory for varieties over finite fields.  This led to the Weil conjectures, the discovery of $\ell$-adic cohomology, etc.  I think this is a pretty good candidate for a powerful but unexpected technique.
A: Gromov's proof that finitely generated groups with polynomial growth are virtually nilpotent. The ingenious step is to consider a scaling limit of the usual metric on the Cayley graph of the finitely generated group. 
Of course the details are messy and to get the final conclusion one has to rely on a lot of deep results on the structure of topological groups. However, already the initial idea is breathtaking.
A: Malliavin's proof of Hormander's theorem is very interesting in the sense that one of the basic ingredients in the language of the proof is a derivative operator with respect to a Gaussian process acting on a Hilbert space.  The adjoint of the derivative operator is known as the divergence operator and with these two definitions one can establish the so called "Malliavin Calculus" which has been used to recover classical probabilistic results as well as give new insight into current research in stochastic processes such as developing a stochastic calculus with respect to fractional Brownian motion.  What makes his proof more interesting is that Malliavin was trained in geometry and only used the language of probability in a somewhat marginal sense at times - alot of his ideas are very geometric in nature which can be seen for example in his very dense book: P. Malliavin: Stochastic Analysis. Grundlehren der Mathematischen Wissenschaften,
313. Springer-Verlag, Berlin, 1997.
A: Barwise compactness and $\alpha$-recursion theory. The idea many properties of the following are captured by thinking of how to define analogs in $V_\omega$: 
(1) Finite sets are elements of $V_{\omega}$.
(2) Computable sets can are $\Delta_1$ definable over $V_{\omega}$. 
(3) Computable enumerable sets can are $\Sigma_1$ definable over $V_{\omega}$. 
(4) First order logic is $L_{\infty, \omega} \cap V_\omega$.
Then, if we replace $V_\omega$ by a different countable admissible set $A$, many of the results relating these classes have analogs. E.g. Barwise compactness, completeness, the existence of an $A$-Turing jump, ... 
A: Lovasz proof of Shannon Capacity of the Pentagon (the only proof known). Introduces Semidefinite optimization. Geometrizes and introduces analytic techniques to Graph Theory. Descartes introduced coordinate space approach to geometric problems. In the same spirit, Lovasz's proof coordinate space approach to graph theory problems.
A: Nicolas Monod's genius two pages new counterexample to the von Neumann conjecture, inspired by Mary Shelley's novel Frankenstein:  http://arxiv.org/abs/1209.5229 
A: I would like to propose the theorem of J.H.C. Whitehead that if $X$ is a path connected space, and $Y$ is formed from $X$ by attaching $2$-cells, i.e. $Y=X \cup_{f_i}e^2_i$ for a family of maps $f_i: S^1 \to X$, then the crossed module $\partial: \pi_2(Y,X,x) \to \pi_1(X,x) \;$ is the free crossed module on the characteristic maps of the $2$-cells. 
The proof spreads over three of his papers. 


*

*On adding relations to homotopy groups. Ann. of Math. (2) 42 (1941) 409--428.

*Note on a previous paper entitled ``On adding relations to  homotopy groups.''.
Ann. of Math. (2) 47 (1946) 806--810.

*Combinatorial homotopy. II. Bull. Amer. Math. Soc. 55 (1949) 453--496.
The essential geometric content of the proof uses transversality and knot theory, and was in his paper 1.  The definition of crossed module was given in his paper 2. Finally the definition of free crossed module was given in his paper 3, together with an outline of the proof, referring back to paper 1. You can find my own exposition of the proof here. The referee wrote that: "The theorem is not new. The proof is not new. But the paper should be published since these papers of Whitehead are notoriously obscure." I explained the proof once to Terry Wall, and he said it was a good 1960's type proof! What my paper does is repackage Whitehead's proof for a modern audience, and with pictures and consistent notation.  
It seems to me pretty good to give the essence of a proof years before you have the right definitions for the theorem! 
The notion of crossed module has over recent years become more widespread, partly because of its relation to $2$-groupoids and double groupoids. This is discussed a little in a  seminar I gave in Chicago last year. See also the Wikipedia entry and that from the nlab. 
A: Quillen's construction of the cotangent complex used homotopical algebra to find the correct higher-categorical object without explicitly building a higher category.  This may sound newfangled and modern, but if you read Grothendieck's book on the cotangent complex, his explicit higher-categorical construction was only able to build a cotangent complex that had its (co)homology truncated to degree 2.  Strangely enough, by the time Grothendieck's book was published, it was already obsolete, as he notes in the preface (he says something about how new work of Quillen (and independently André) had made his construction (which is substantially more complicated) essentially obsolete).  
A: I find Shannon's use of random codes to understand channel capacity very striking.  It seems to be very difficult to explicitly construct a code which achieves the channel capacity - but picking one at random works very well, provided one chooses the right underlying measure.  Furthermore, this technique works very well for many related problems.  I don't know the details of your Example 4 (Erdos and Ramsey numbers), but I expect this is probably closely related.
A: Emil Artin's solution of Hilbert's 17th problem which asked whether every positive polynomial in any number of variables is a sum of squares of rational functions.
Artin's proof goes roughly as follows. If $p \in \mathbb R[x_1,\dots,x_n]$ it not a sum of squares of rational functions, then there is some real-algebraically closed extension $L$ of the field of rational functions in which $p$ is negative with respect to some total ordering (compatible with the field operations), i.e. there exists a $L$-point of $R[x_1,\dots,x_n]$ at which $p$ is negative. However, using a model theoretic argument, since $\mathbb R$ is also a real-closed field with a total ordering, there also has to be a real point such that $p<0$, i.e. there exists $x \in \mathbb R^n$ such that $p(x)< 0$. Hence, if $p$ is everywhere positive, then it is a sum of squares of rational functions.
The ingenius part is the use of a model theoretic argument and the bravery to consider a totally ordered real-algebraic closed extension of the field of rational functions. 
A: And how about Perelman's proof of Poincare's conjecture?
A: Heegner's solution to the Gauss class number 1 problem for imaginary quadratic fields, by 
noting that when the class number is 1 then a certain elliptic curve is defined over Q and
certain modular functions take integer values at certain quadratic irrationalities, and then
finding all the solutions to Diophantine equations that result, seems to me equally beautiful
and unexpected. Maybe its unexpectedness kept people from believing it for a long time.
A: Sometimes mathematics is not only about the methods of the proof, it is about the statement of the proof.  E.g., it is hard to imagine an theorem-searching algorithm ever finding a proof of the results in Shannon's 1948 Mathematical Theory of Communication, without that algorithm first "imagining" (by some unspecified process) that there could BE a theory of communication.
Even so celebrated a mathematician as J. L. Doob at first had trouble grasping that Shannon's reasoning was mathematical in nature, writing in his AMS review (MR0026286):[Shannon's] discussion is suggestive throughout, rather than mathematical, and it is not always clear that the author's mathematical intentions are honorable.The decision of which mathematical intentions are to be accepted as "honorable" (in Doob's phrase) is perhaps very difficult to formalize.


[added reference] 
One finds this same idea expressed in von Neumann's 1948 essay The Mathematician:Some of the best inspirations of modern mathematics (I believe, the best ones) clearly originated in the natural sciences. ... As any mathematical discipline travels far from its empirical source, or still more, if it is a second or third generation only indirectly inspired by ideas coming from "reality", it is beset by very grave dangers. It becomes more and more purely aestheticizing, more and more l`art pour le art. ... Whenever this stage is reached, the only remedy seems to me to be the rejuvenating return to the source: the reinjection of more or less directly empirical ideas.One encounters this theme of inspiration from reality over-and-over in von Neumann's own work.  How could a computer conceive theorems in game theory ... without having empirically played games?   How could a computer conceive the theory of shock waves ... without having empirically encountered the intimate union of dynamics and thermodynamics that makes shock wave theory possible?  How could a computer conceive theorems relating to computational complexity ... without having empirically grappled with complex computations?

The point is straight from Wittgenstein and E. O. Wilson: in order to conceive mathematical theorems that are interesting to humans, a computer would have to live a life similar to an ordinary human life, as a source of inspiration.
A: The use of ideals in rings, rather than elements (in terms of factorization, etc...).
This was followed by another revolutionary idea: using radical (Jacobson radical, etc...) instead of simple properties on elements.
A: Grothendieck's insight how to deal with the problem that whatever topology you define on varieties over finite fields, you never seem to get enough open sets. You simply have to re-define what is meant by a topology, allowing open sets not to be subsets of your space but to be covers. 
I think this fits the bill of "seem very natural once you are used to it", but it was an amazing insight, and totally fundamental in the proof of the Weil conjectures. 
A: Shigefumi Mori's proof of Hartshorne's conjecture (the projective spaces are the only smooth projective varieties with ample tangent bundles). In his proof, Mori developed many new techniques (e.g. the bend-and-break lemma), which later became fundamental in birational geometry.
A: Morse theory is another good example. Indeed it is the inspiration for Floer theory, which has already been mentioned.
Atiyah-Bott's paper "Yang-Mills equations on a Riemann surface" and Hitchin's paper "Self-duality equations on a Riemann surface" both contain rather striking applications of Morse theory. The former paper contains for example many computations about cohomology rings of moduli spaces of holomorphic vector bundles over Riemann surfaces; the latter paper proves for instance that moduli spaces of Higgs bundles over Riemann surfaces are hyperkähler.
Note that these moduli spaces are algebraic varieties and can be (and are) studied purely from the viewpoint of algebraic geometry. But if we look at things from an analytic point of view, and we realize these moduli spaces as quotients of infinite dimensional spaces by infinite dimensional groups, and we use the tools of analysis and Morse theory, as well as ideas from physics(!!!), then we can discover perhaps more about these spaces than if we viewed them just algebraically, as simply being algebraic varieties.
A: Do Cantor's diagonal arguments fit here?  (Never mind whether someone did some of them before Cantor; that's a separate question.)
A: Lobachevsky and Bolyai certainly introduced a fundamentally new way of thinking, though I'm not sure it fits the criterion of being a proof of something - perhaps a proof that a lot of effort had been wasted in trying to prove the parallel postulate. 
A: Generating functions seem old hat to those who have worked with them, but I think their early use could be another example.  If you did not have that tool handy, could you create it?
Similarly, any technique that has been developed and is now widely used is made to look natural after years of refining and changing the collected perspective, but might it not have seemed quite revolutionary when first introduced?  Perhaps the question should also be about such techniques.
Gerhard "Old Wheels Made New Again" Paseman, 2010.12.09
A: I am surprised that noone mentioned Hilbert's proof of Hilbert's Basis Theorem yet. It says that every ideal in $\mathbb{C}[x_1,\ldots,x_n]$ is finitely generated - the proof is nonconstructive in the sense that it does not give an explicit set of generators of an ideal. When P. Gordan (a leading algebraists at that time) first saw Hilbert's proof, he said, "This is not Mathematics, but theology!"
However, in 1899, Gordan published a simplified proof of Hilbert's theorem and commented with "I have convinced myself that theology also has its advantages."
A: I think Fürstenberg´s proof of the infinitude of primes, taken for itself, could be considered in the spirit of the original question, even though its mathematical value is questionable.
A: Bourgain,following Gower's ideas on Balog-szemeredi gave sharpest bound till then for n>8  on minkowski dimension in kakeya problem. the work used ideas from arithmetic combinatorics to harmonic analysis and charles fefferman comments that it gives the feeling from where in Mars did it come from.
A: Dvir's proof of the finite field Kakeya conjecture. 
I know this is closely related to the Guth--Katz result which has already been dismissed, but I will justify why it answers the question. Dvir's proof dramatically changed the point of view in the restriction problem (a big subfield of Euclidean harmonic analysis) and allied problems in PDEs and geometric meaure theory by importing 'the polynomial method' into the field. In particular, it is highly unlikely that the Guth--Katz result would have came as and when it did without Dvir's insight. Prior to Dvir's insight a surprisingly large amount of effort was spent on working towards the finite field Kakeya conjecture (Dvir's theorem) including many works by a Field's medalist. While Dvir's original proof was a few pages, it was quickly distilled into a paragraph proof that is accesible to undergraduates who have just understood what a finite field is. 
The obvious critique is that the polynomial method existed before this. Certainly it did and polynomials for a long time before that. I am not saying that Dvir invented the method. However, no one saw that this connection nor expected it. Similarly complex analysis and number theory were investigated in parallel before Riemann introduced his famous eponymous hypothesis. In analogy with the above example on the infinitude of primes, it would be like we started with Euler's proof, then discovered Hadamard--de la Vallee Poussin's proof with the twist that their proof was only as transparent and elegant as Euclid's proof. 
A: The post,the answers generated, the comments...wow! They are so nice! 
Though I am not going to answer in exactly the way required, I believe including occasions were new insights helped to give support a great unsolved problem are worth-noting. 
For instance, we can consider a case in Random Matrix Theory: the statistical interplay between the distribution the zeros of the Riemann Zeta function and the eigenvalues of a random Hermitian matrix which has provided a basis for the Hilbert–Pólya conjecture. .
A: Hilbert's proof of Hilbert-Waring theorem.
A: The first formal proofs using limits. (the oldest ones I know are in Newton's Principia)
A: I don’t believe that Hilbert’s 10th problem has been mentioned ... all of what went into that.
I also don’t believe that all of what went into making it possible to prove the four color theorem with the help of a computer has been mentioned.
And although much is mentioned about set theory, introducing the actual axioms of ZFC was revolutionary and I don’t think it has been mentioned ... apologies if I missed it in the long set of replies ... I guess this answer can always be edited (or deleted if it is a duplicate).
