The best coding AI in the world right now is not on the SWE-bench public leaderboard the way you’re used to reading it. It tops the leaderboard at 93.9% on SWE-bench Verified, a number that, three weeks after it landed, is still the highest score any model has posted on that benchmark. It also has roughly 52 customers. You and your team are almost certainly not among them, and according to Anthropic’s own statements, that’s the point.
Claude Mythos Preview was unveiled on April 7, 2026, alongside an unusual companion announcement called Project Glasswing. The model is technically the strongest thing Anthropic has ever shipped. The shipping list is twelve named launch partners and forty-odd additional organizations that maintain critical infrastructure. Everyone else — the open developer market that Claude has spent four years winning — is locked out, on purpose, indefinitely.
This is new. Capability gating at the frontier — a lab choosing not to ship the strongest model it has — was an abstract policy question six months ago. As of late April 2026, it is the actual product strategy of the lab that, by most independent measures, has been ahead on coding for three consecutive releases. It deserves a closer look than it’s getting.
What Mythos Actually Did on the Benchmarks
The headline number is the SWE-bench Verified score. SWE-bench Verified is a curated set of real GitHub issues from real Python repositories — Django, Flask, scikit-learn, pytest — where each task gives a model a codebase, a bug report, and the unit tests that define “fixed.” In mid-2024, the best agent frameworks scored in the 40–55% range. Human engineers, given a reasonable time budget per task, score around 67–70% on independent runs. Mythos Preview scores 93.9%. That is approximately nineteen real bugs out of twenty resolved correctly against unit tests written by other humans.
The rest of the spec sheet is similarly steep relative to the most recent generally-available Anthropic flagship, Claude Opus 4.6:
| Benchmark | Claude Mythos Preview | Claude Opus 4.6 (prior flagship) |
|---|---|---|
| SWE-bench Verified (real coding) | 93.9% | 80.8% |
| SWE-bench Pro (harder, recent variant) | 77.8% | 53.4% |
| Terminal-Bench 2.0 (agentic) | 82.0% | 65.4% |
| USAMO 2026 (math olympiad) | 97.6% | 42.3% |
| Multimodal (MMMU class) | ~59% | ~57% |
The pattern in that table is the most interesting thing about Mythos. The jumps in coding, agentic, and mathematical reasoning are enormous — SWE-bench Pro alone moved 24 percentage points in one generation. The jump in multimodal is essentially a rounding error. Mythos is not a uniformly stronger Claude. It is a coding-and-reasoning specialist that happens to be the most capable model Anthropic has ever benchmarked, and it’s a reminder that the “no single best AI” thesis we’ve been writing about since the three-flagship month doesn’t go away when capabilities scale — it sharpens.
Why “You Can’t Use It” Is the Product, Not the Bug
Anthropic’s framing for the gating is unusually direct. From Project Glasswing’s announcement: AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities. Mythos Preview has already discovered thousands of high-severity vulnerabilities, including some in every major operating system and every major web browser.
The same capability that makes a model 93.9% on SWE-bench Verified makes it a world-class autonomous vulnerability finder. Those are not two skills. They are one skill, applied to two prompts.
That’s the conceptual core of the gate. SWE-bench rewards a model for reading a real repository, building a mental model of it, locating the lines that matter, and patching them. Vulnerability discovery rewards a model for reading a real repository, building a mental model of it, locating the lines that matter, and exploiting them. Read the verbs side by side and the symmetry is obvious. Anthropic’s position is that the only reasonable way to ship the upside without immediately handing the downside to every script kiddie on the internet is to put the model in the hands of defenders first and everyone else later, if at all.
The Glasswing launch partners are unusually well-chosen if that’s the actual goal: AWS, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. That list maps almost perfectly onto the surface area of the world’s critical software — the major clouds, the major operating system vendors, the major endpoint and network security vendors, the dominant open-source steward, and one large-bank representative of the regulated-industry side. Roughly forty additional organizations get access on top of the twelve launch partners. Anthropic has put up $100 million in Mythos usage credits and $4 million in direct donations to open-source security organizations to make sure the under-resourced part of that surface area can play. They have also been in ongoing conversations with US government officials about offensive and defensive cyber implications.
You can argue with the policy. You can argue Anthropic should have published more, gated less, or both. What you can’t argue is that this is incoherent. It is the first time a frontier lab has deliberately decided that the strongest model it can build is too strong to put on a public API, and built a defensible institutional structure around the decision in public.
What Capability Gating Means for the Rest of Us
For the developer who isn’t at one of those 52 organizations, three things change in practical, day-to-day terms.
1. The public leaderboard now has an invisible ceiling
For the last two years, the public SWE-bench Verified leaderboard was a reasonable proxy for “what is the best coding AI you can buy.” It is no longer that. As of late April 2026, the public top of SWE-bench Verified is Claude Opus 4.7 (Adaptive) at 87.6%, with GPT-5.5 / Codex variants close behind at roughly 85%. The actual frontier sits at 93.9%, but the actual frontier isn’t for sale. Anyone who tells you in May 2026 that they’re using “the best AI for coding” is using the best one available on a credit card — not the best one in existence. Worth keeping straight.
2. Multi-model comparison is now the only way to claw back some of that gap
The only way to approximate Mythos-class reliability without Mythos is to run the strongest publicly-available models in parallel and exploit the geometry of where they agree. Empirically, when Opus 4.7 (87.6%) and GPT-5.5 (~85%) converge on the same patch for a real GitHub issue, the conditional accuracy of that patch sits comfortably above either individual score. When they diverge, you’ve found exactly the issues to look at by hand. This is the same logic SNEOS Premium’s Trust Scoring uses on consumer prompts, scaled down to a smaller panel: agreement is signal, disagreement is a flashlight.
3. The “no single best AI” thesis isn’t weakened by Mythos. It’s reinforced.
It would be tempting to read “model X scores 93.9% on the most-watched coding benchmark” as a final answer to the “which AI is best?” question. But re-read the Mythos benchmark table. It dominates coding, agentic, and math. It barely moves on multimodal. There is no scenario in May 2026 where you build, say, a video-understanding workflow on Mythos — and even at 52 customers, Glasswing partners are routing video and image work through other models. Specialization is now a permanent property of the frontier, not a transitional one. The strongest model on the planet still isn’t the strongest at everything.
The New Three-Tier Map of the AI Landscape
Mythos crystallizes a three-tier structure that has been quietly forming for six months. As of May 1, 2026 it’s clear enough to put in a chart:
| Tier | Examples | Who can use it |
|---|---|---|
| Locked-vault frontier | Claude Mythos Preview | ~52 vetted Project Glasswing partners |
| Public frontier | Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro | Any team with an API key and a budget |
| Open-weight chasers | DeepSeek V4, Llama 4 Behemoth (expected), Gemma 4 | Any team that wants to self-host |
Three years ago all three tiers were the same tier. Eighteen months ago the locked-vault row was theoretical. Today it has its first real entry, and there is no good reason to expect it to stop at one. If Mythos genuinely does what Anthropic says it does on vulnerability discovery, OpenAI and Google are now under pressure either to gate their own next-gen coding-class models or to publicly explain why they aren’t. Neither answer leaves the rest of us in the same world we were in last year.
How to Route in a World Where the Best Model Isn’t For Sale
The practical question on May 1 is: knowing all of the above, what should an actual team do this week? A short version of the playbook:
- For coding-heavy work, route to Claude Opus 4.7 first, double-check with GPT-5.5. If they agree on the patch, ship it. If they don’t, the disagreement is the signal.
- For long-running agents and terminal automation, GPT-5.5 first, Opus 4.7 second. The Terminal-Bench gap between Mythos (82%) and the public frontier is real, but smaller than the SWE-bench gap, so good multi-model consensus closes most of it.
- For multimodal work, Mythos genuinely doesn’t matter. Gemini 3.1 Pro remains the right starting point, with GPT-5.5 as a second opinion.
- For high-stakes regulated content (legal, medical, financial), run all three flagships and treat the consensus as your primary signal — the same approach we wrote up in our legal contract review story still holds, and it holds harder now that the actual frontier is invisible.
- For self-hosted or cost-sensitive workloads, the open-weight tier (DeepSeek V4, Gemma 4, the Llama 4 Behemoth release expected this month) is where the leaderboard fight is most active and where prices are dropping fastest.
Notice that none of these answers is “just pick one model.” In a world where the actual frontier is gated, the rational move for everyone outside the gate is to treat the public frontier as a panel, not a pick. That’s the entire reason SNEOS exists in the form it does — one prompt, all the models you’re actually allowed to use, side by side, with Trust Scoring on Premium doing the agreement-geometry math automatically.
The Precedent Is Bigger Than the Model
Three weeks in, Mythos itself is interesting. The precedent it sets is more interesting. For most of the last decade, the implicit deal between frontier AI labs and the rest of the software industry was: when we build it, you’ll be able to buy it. That deal still holds for Opus 4.7, for GPT-5.5, for Gemini 3.1 Pro — the public frontier is genuinely excellent, and getting cheaper every quarter. But the implicit deal no longer holds for the absolute frontier. The absolute frontier, at least at one lab, now ships to defenders first. Whether other labs follow is the live question for the rest of 2026.
If they do, the “which AI should I use?” question splits permanently into two questions: which gated model is your organization important enough to access, and which public models are you running in parallel to approximate the rest. Those are different questions. Most teams will only ever have to think about the second. But pretending the first one doesn’t exist, or that the public leaderboard is still the whole picture, is no longer accurate as of April 7, 2026.
The most interesting AI release of 2026 is the one you can’t use. The right response isn’t to wait for it — Anthropic has been clear that public release is not currently planned. The right response is to do, with the models you can use, the kind of careful comparison that approximates what Mythos would do alone. That is, in one sentence, the entire SNEOS thesis. Mythos didn’t change it. Mythos confirmed it.
Are you on a Glasswing partner team? We’d love to hear (off the record) how Mythos is changing your security workflow. Drop us a line.