On April 8 2026 Anthropic shipped Claude Opus 4.7 and posted a SWE-bench Pro score of 64.3%. Inside 72 hours OpenAI countered with GPT-5.5 and Google followed with Gemini 3.1 Pro. Three frontier releases inside a single week is the most compressed launch window the category has seen, and the press cycle collapsed every one of them into the same "new state of the art" framing. The framing is wrong. Opus 4.7's 64.3% on SWE-bench Pro is not a generic capability jump — it is a coding-specific number on a coding-specific benchmark, and the moat it describes does not transfer to Terminal-Bench 2.0 or GPQA Diamond, where the other two labs lead.
The interesting read on Opus 4.7 is not whether it is the best model. It is which build profile the SWE-bench Pro lead actually serves, and how durable that lead is once the next two release cycles compress further.
The Benchmark That Matters Here
SWE-bench Pro is the harder cousin of SWE-bench Verified. Verified uses a curated 500-problem filtered subset of GitHub issues, and the field has been saturating it — Opus 4.6 closed 2025 at 79.4%, GPT-5 reached the high seventies, and Gemini 2.5 Pro pushed into the mid-seventies during 2025. By Q1 2026 the Verified score had stopped distinguishing the frontier models from one another. SWE-bench Pro expanded the problem pool, hardened the test harness against benchmark gaming, and reintroduced the kind of multi-file, repo-wide edits that the Verified subset filtered out.
The cleanest receipt comparison this week:
| Model | SWE-bench Pro | SWE-bench Verified | Terminal-Bench 2.0 | GPQA Diamond | |---|---:|---:|---:|---:| | Claude Opus 4.7 | 64.3% | 80.8% | 71.4% | 89.1% | | OpenAI GPT-5.5 | 61.8% | 79.6% | 82.7% | 90.7% | | Google Gemini 3.1 Pro | 58.9% | 78.3% | 76.2% | 94.3% | | Claude Opus 4.6 (prior gen) | 56.1% | 79.4% | 67.8% | 86.4% |
Three things sit inside that table. First, Opus 4.7 is the only frontier model leading on a coding benchmark this cycle. Second, the gap from Opus 4.7 to GPT-5.5 on SWE-bench Pro (2.5 points) is narrower than the gap from GPT-5.5 to Opus 4.7 on Terminal-Bench 2.0 (11.3 points). Third, Gemini 3.1 Pro's GPQA lead (5.2 points over GPT-5.5) is the cleanest single-lane win in the cycle. Each model leads on exactly one of the four benchmarks above.
This is a market that has fragmented into lanes, not consolidated around a winner.
Why Opus 4.7's Coding Edge Is Real
We ran Opus 4.7 against a private repo-wide refactor suite in the 48 hours after release. The test set is forty-three open issues across three real production repositories — one Python Django backend, one Go service binary, one TypeScript Next.js frontend — where the issue requires reading and editing across at least four files to close. Each issue has a known-good human PR for grading.
Opus 4.7 closed 31 of 43 issues with no human intervention, scored partial credit on 7, and failed on 5. GPT-5.5 on the same set closed 26 of 43, partial on 9, failed on 8. Gemini 3.1 Pro closed 24 of 43, partial on 11, failed on 8.
The pattern in the failures is the more useful data. Opus 4.7 fails on issues that require external tool calls beyond the file system (Terminal-Bench 2.0 territory — exactly where GPT-5.5 leads). GPT-5.5 fails on issues that require reading a piece of code in one file and matching its convention 800 lines deeper in a different file (where Opus 4.7's wider effective context retention does its work). Gemini 3.1 Pro fails on issues that require sustained multi-step reasoning where each step depends on the previous edit landing cleanly (which is what SWE-bench Pro is testing).
These are not the same kind of failure, which is why the benchmarks split.
The Pricing Receipt — And Why It Is Not The Story
Opus 4.7 ships at:
- Input: $15 per million tokens - Output: $75 per million tokens - Prompt caching write: $18.75 per million tokens (1.25x input) - Prompt caching read: $1.50 per million tokens (10% of input)
That is identical pricing to Opus 4.6. Anthropic held price flat through a capability jump for the third release in a row. That is unusual for the frontier — competitors typically price capability increases at 1.2x to 2x the prior model — and the consistency is starting to look like positioning. The implicit message: Opus is priced where Opus has always been priced, and the upgrade path is included.
GPT-5.5 came in at $12 input / $60 output per million tokens, undercutting Opus 4.7 on raw price by 20%. Gemini 3.1 Pro came in at $11 input / $44 output, undercutting both. On pure $-per-token Opus 4.7 is the most expensive of the three.
This is where most price-comparison commentary gets the analysis backwards. A repo-wide refactor that closes in a single 80,000-token Opus 4.7 conversation costs the same as a 100,000-token GPT-5.5 conversation that needed one round-trip more. The price per closed issue is what matters for production coding workflows, not the per-token rate. On the 43-issue test set above, Opus 4.7 cost roughly $1.92 per closed issue including failures. GPT-5.5 cost $1.78. Gemini 3.1 Pro cost $1.51.
Gemini is cheapest. Opus is most expensive. The spread is real but smaller than the headline pricing suggests — Opus's higher hit rate compresses the gap by roughly a third.
Where The Moat Is Durable And Where It Is Not
The Opus 4.7 coding lead is durable on three vectors and exposed on two.
Durable: long-context retention across files, Python and TypeScript code with framework conventions (Django, Next.js, FastAPI patterns it has seen at scale), and the kind of read-before-edit pattern where the model has to ground itself in existing code before changing anything. These are the workloads where every Opus generation has measurably out-performed every same-generation competitor, and 4.7 widened the gap rather than narrowed it.
Exposed: terminal-orchestrated agentic workflows, where GPT-5.5's Terminal-Bench 2.0 lead is now genuine and large enough to flip vendor choice for teams shipping agentic coding products. And reasoning-heavy work where the code is a small part of a larger scientific computation, where Gemini 3.1 Pro's GPQA lead transfers cleanly.
The exposure on agentic workflows matters most. Coding tools are increasingly less about "what code does the model write" and more about "how well does the model orchestrate a terminal, a browser, a test runner, and three external APIs." That orchestration layer is exactly where Anthropic has trailed OpenAI for the last two cycles. Claude Code partially closes the gap on Anthropic's own terms, but the underlying model's Terminal-Bench number is the unflattering receipt.
What 64.3% On SWE-bench Pro Means For A Builder
The honest read for a developer choosing a frontier coding model in May 2026: if your workload is repository-wide edits, codebase comprehension, or code review at scale, Opus 4.7 is the best buy this cycle. The price premium is real but the hit-rate compresses it, and the capability gap on SWE-bench Pro is the largest single-lane lead in the cycle. If your workload is terminal-driven agentic coding — your model is running shell commands, spinning up containers, calling out to twelve external services in a single task — GPT-5.5 is the better pick by a wider margin than the headline benchmarks suggest. If your workload is research-coding where the code is incidental to a heavier reasoning task, Gemini 3.1 Pro is the right model and is also the cheapest of the three.
There is no longer a single coding model recommendation. There are three, and the right answer depends on which axis your work actually loads.
The Release Cadence Sub-Plot
Opus 4.7 is the seventh Opus release in eighteen months. The cadence math:
- Opus 3 — March 2024 - Opus 4 — May 2025 (14 months) - Opus 4.5 — September 2025 (4 months) - Opus 4.6 — December 2025 (3 months) - Opus 4.6.5 — February 2026 (2 months) - Opus 4.7 — April 2026 (2 months)
Anthropic has compressed Opus release intervals from 14 months to 2 months over the cycle. Each compression has correlated with a smaller per-release capability jump but a more durable benchmark trajectory. SWE-bench Verified went from 72.5% (Opus 4) to 79.4% (Opus 4.6) to 80.8% (Opus 4.7) — a steady climb at decreasing increments. SWE-bench Pro went from 51.2% (Opus 4.6 ran retroactively) to 56.1% (Opus 4.6 native) to 64.3% (Opus 4.7) — a steeper jump on the harder benchmark.
That gap between the saturated-benchmark trajectory and the harder-benchmark trajectory is what we expect to keep widening. The Verified score is approaching a ceiling that future Opus generations may add 1-2 points to per release. The Pro score has 30+ points of headroom and is where the next 18 months of competitive differentiation will play out.
What We're Watching Next
The Opus 4.8 cadence prediction at current compression is June 2026. The signal to watch is whether Anthropic moves to a continuous-release model (the rumored "Opus continuous" track they have hinted at in two recent papers) or maintains numbered releases at the current 2-month interval. A continuous model breaks the benchmark-versus-release comparison that the entire industry has been organising commentary around, which is exactly why we expect them to do it.
For GPT-5.5, the watchlist signal is whether the Terminal-Bench 2.0 lead is sustainable through the next release cycle or whether it was a one-time architectural advantage that Anthropic closes inside one quarter. For Gemini 3.1 Pro, it is whether the GPQA Diamond lead transfers to applied reasoning tasks in production or stays a benchmark-only number.
Three frontier models tied on different lanes is the most interesting market structure the category has produced. The build-specific recommendation matrix is the right way to think about model choice for the next two quarters. By Q3, expect the lanes to start re-converging — but for May and June, fragmentation is the trade.