Claude Opus 4.7: 64.3% on SWE-bench, 30% cost cut at Box, and Mythos still out of reach
Alexandre
··
Reading time: 8 min
Anthropic released Claude Opus 4.7 on April 16, 2026. 64.3% on SWE-bench Pro (vs 57.7% for GPT-5.4 and 53.4% for Opus 4.6). Box cut their AI infrastructure costs by 30%. XBOW jumped from 54.5% to 98.5% on autonomous penetration testing. And in that same benchmark table, Anthropic showed a model that's even better... that nobody can access.
Opus 4.7 dropped four days after the Opus 4.6 crisis, where thousands of devs had documented regressions on GitHub, Hacker News, and X. The timing is not a coincidence.
Here's the breakdown.
At a glance
Opus 4.7 leads on agentic coding (a model's ability to resolve real bugs on open-source repos, autonomously) among models available in April 2026. Mythos Preview dominates across the board, but you can't use it.
Mythos Preview isn't in this table because it isn't generally available. Anthropic published its numbers anyway. More on that below.
What actually changed
Opus 4.7 fixes the three main pain points of Opus 4.6: it checks its outputs before reporting back, it no longer skips step 4 of 6 in an instruction chain, and it holds coherence across multi-hour sessions. Image resolution goes from 685 to 2,576 pixels, and a new xhigh effort level replaces the previous maximum.
Self-verification. The model reviews what it produces before reporting back. Tool errors dropped to a third of the prior level, confirmed by both Anthropic and enterprise testers (aimlapi.com). If you have ever seen Claude say "done" when 80% of the work was still left, you know exactly what this fixes.
Instruction tracking. The thing that drove people crazy on 4.6 was multi-step chains where it cut a corner in the middle. Give it 6 steps, it does 3 correctly and bungles the rest. On 4.7, multi-step agentic reasoning improved by 14% according to early access testers. In practice, it no longer skips steps.
Long-context coherence. Over a multi-hour session, the first instruction survives. That's exactly what Opus 4.6 broke starting in February: instruction drift, where the model forgets what you told it in hour one by hour two. Devs have still flagged regressions in specific edge cases, so it is not perfect everywhere.
3.75x image resolution. Now at 2,576 pixels natively, compared to around 685 on 4.6. CharXiv Reasoning reaches 91.0% with tools (up from 84.7%). For anyone working with UI screenshots, architecture diagrams, or technical documents, this is a real change: no more pre-processing images just so the model can read them (ayautomate.com).
xhigh effort level. One notch above high. More latency, deeper reasoning. It is the default in Claude Code with Opus 4.7.
The reaction on X
Two camps. Those who see Opus 4.7 as rehabilitation after the 4.6 crisis. And those wondering why we are celebrating a model when Anthropic is publicly showing it has a better one under wraps.
The official announcement from @claudeai:
Chargement du tweet...
And from the developer-focused account @ClaudeDevs:
Chargement du tweet...
Devs frustrated with 4.6 since February took this as vindication. An AMD senior director had written on GitHub: "Claude has regressed to the point it cannot be used for complex engineering." That post circulated widely. April 16's numbers are a direct answer to that.
One dev noted something practical: "Opus 4.7 is in Claude Code, the new xhigh effort level is the default now." The skeptical take followed immediately: "My guess is it is dynamic thinking that uses fewer tokens." Possible. The other camp is not celebrating anything. If Mythos Preview already exists and is already better, what exactly are we celebrating?
The Gizmodo headline sums up the mood: "Anthropic Releases Claude Opus 4.7 to Remind Everyone How Great Mythos Is."
The numbers from actual companies
Box cut model calls by 56%, tool calls by 50%, and its AI bill by 30% switching to Opus 4.7, at equivalent output quality. XBOW went from 54.5% to 98.5% on autonomous pen-testing. That is field data, not a lab.
Yashodha Bhavnani, Head of AI at Box, published the numbers: 56% fewer model calls, 50% fewer tool calls, 24% faster response times, and 30% fewer AI units consumed compared to Opus 4.6, at the same output quality. For a company running this at scale, that is a concrete infrastructure story.
Replit got the same code quality at lower cost. Consistent signal.
XBOW, the autonomous pen-testing company, jumped from 54.5% to 98.5%. The biggest single-model jump I have seen on any model change. Anthropic notes they deliberately limited Opus 4.7's cyber capabilities compared to Mythos Preview (new safeguards are tested on the less powerful model first). The number comes with that caveat. It is still 98.5%.
One fintech platform in early access summed up self-verification in practice: "It catches its own logical faults during the planning phase and accelerates execution, far beyond previous Claude models."
Personally, the Box number speaks to me most. 30% cost reduction at equal quality is the kind of thing that justifies migrating a production pipeline without stress. On Waku with Claude Code, fewer tool calls translates directly into cleaner sessions. Fewer retries, fewer pointless loops, fewer tokens burned for nothing.
The hidden cost: the tokenizer inflates the bill by 5 to 35%
The listed price has not changed: $5 per million input tokens, $25 per million output tokens. But Opus 4.7's new tokenizer inflates token counts by 5 to 35% depending on the use case. You pay the same per token, but consume more tokens for the same thing.
This does not show up anywhere in the official announcement. The sticker price is identical to Opus 4.6. But the tokenizer changed, and the same conversation, the same prompt, the same code file produces more tokens than before (ayautomate.com). On light usage, it is invisible. On an automated pipeline processing thousands of files per day, it shows up on the bill.
So before migrating: test your heaviest pipelines (vision, long documents, full repos) and compare token counts before and after. Prompt caching reduces up to 90% of input costs, and the Batch API offers 50% off. But those optimizations already existed on 4.6. The real delta is the tokenizer.
I documented this kind of trap in my guide to cutting token consumption in half. The savings do not come at the model level, but at the workflow level: what you send, how you structure your prompts, and how much context you load for nothing.
Mythos: the ceiling you can see but cannot reach
Anthropic released Opus 4.7 with a benchmark table that includes Mythos Preview, a model that is better on every dimension, which you cannot access. SWE-bench Pro: Mythos at 77.8%, Opus 4.7 at 64.3%. Terminal coding: 82.0% vs 69.4%. Vulnerability reproduction: 83.1% vs 73.1%.
Mythos Preview launched on April 7, 2026 under Project Glasswing, initially accessible to 11 companies for cybersecurity vulnerability discovery. Then expanded to around 150 partners across more than fifteen countries. No general availability. Apple and some major platform vendors are among the first.
In June, things got complicated. On June 12, the US Department of Commerce imposed a suspension of all access to Mythos-class models for foreign nationals, including Anthropic employees (Wikipedia). On June 26, access was restored for more than 100 US organizations, but Mythos remains inaccessible for non-Americans. Anthropic is preparing a "trusted-access" program for defensive cybersecurity and biomedical research. No timeline for general availability.
So yes, Opus 4.7 is the best model available for most devs. It leads on agentic coding among public models. But Anthropic keeps publishing benchmarks for a model you cannot use, right next to the one you are paying for. Honestly, that is a comms choice that creates more frustration than hype.
The real test is not launch day. It is three weeks later, when you are working with it without thinking about it. On Waku, the instruction drift that made Opus 4.6 painful past the second hour has not come back once.
My personal data point. Box's 30% cost reduction is the enterprise signal. Mine: I stopped second-guessing the model mid-session.
Not because it produces more. Because it is more predictable. When you know the session is going to hold, you plan differently. You load heavier tasks. You trust the first result instead of re-running three times to check. My full setup is in my Claude Code post, and the architecture of AI code agents explains why coherence over time actually matters when you are working solo.
Since then, Opus 4.8 shipped with a reworked adaptive thinking. The landscape keeps moving. Opus 4.7 is the model that rebuilt trust after the 4.6 crisis.
3 things to remember
1. Predictability beats raw power
Box's 30% savings and the +14% improvement in multi-step reasoning matter. But the real gain is not second-guessing mid-session. A model you can rely on for three hours changes how you plan your day.
2. The tokenizer is a hidden cost
Same listed price, 5 to 35% more tokens. On solo use, it is marginal. On an automated pipeline, it is a line item that creeps up without warning. Test before migrating.
3. Mythos is a signal, not a product
Publishing a better and inaccessible model is a PR move. In practice, Opus 4.7 (then 4.8) is what we ship with. The real issue is the reliability of the model you use every day, not benchmarks for something you cannot touch.
Conclusion
Opus 4.7 does what Opus 4.6 should have done from the start: read before writing, check before reporting, hold instructions over time. Enterprise numbers (Box, XBOW, Replit) confirm the gain. The tokenizer quietly inflates the bill. And Mythos remains a visible but unreachable ceiling.
The tool I am shipping with tomorrow is this one. Have you felt the difference between 4.6 and 4.7 on your projects? Send me a message on X or in the comments.
Alex
Key takeaways
Opus 4.7 leads on agentic coding with 64.3% on SWE-bench Pro, ahead of GPT-5.4 (57.7%) and Opus 4.6 (53.4%). Self-verification and instruction tracking fix the main pain points from the previous version.
Box cut AI costs by 30% and model calls by 56% switching to Opus 4.7, at equivalent quality. XBOW went from 54.5% to 98.5% on autonomous pen-testing.
The tokenizer inflates token counts by 5 to 35% despite identical listed prices. Test your pipelines before migrating.
Mythos Preview dominates every benchmark but remains inaccessible. Access was suspended in June 2026 for non-Americans following US Department of Commerce export controls.
Predictability beats raw power. Opus 4.7 rebuilt the trust that Opus 4.6 had broken.
