Anthropic shipped Claude Opus 4.7 on April 16. It costs the same as Opus 4.6. The context window stays at one million tokens. Platform availability – API, Claude.ai, AWS Bedrock, Google Cloud Vertex AI, Microsoft Foundry – is unchanged in structure. Those facts are confirmed.
The performance claims are a different story. Three benchmark figures are circulating: a 3x production task improvement on Rakuten-SWE-Bench, a 13% gain on an internal 93-task coding evaluation, and a 70% score on CursorBench versus 58% for Opus 4.6. Each of those numbers carries a different level of verifiability. For developers evaluating whether to migrate a production agentic pipeline, that distinction matters more than the numbers themselves.
Section 1: What Actually Changed
Three capability areas have enough source grounding to describe with confidence, though two of the three still lack confirmed documentation URLs.
xhigh effort level and task budgets. Anthropic’s release announcement introduces an “xhigh” effort level sitting between the existing “high” and “max” settings. For agentic loops where token spend scales with task complexity, this gives developers a new control point. Task budgets, currently in beta, let you cap token expenditure per loop iteration. The practical use case is direct: long-running autonomous coding agents can now be cost-capped without forcing a binary choice between “high” and the most expensive reasoning tier. The technical documentation for these features lives at Anthropic’s developer portal; the specific docs URL is pending resolution through the pipeline’s Anthropic API documentation.
Vision resolution increase. Anthropic states the maximum supported image resolution increases to 2,576px (approximately 3.75 megapixels), up from the prior approximately 1,568px limit. The pixel math checks out, a 2,576 × 1,456 image produces roughly 3.75 megapixels. For computer-use and UI automation tasks, more visual resolution means the model can parse denser interfaces, smaller text, and more complex screen states before requiring a crop or downscale step. This is a meaningful operational upgrade for teams building desktop automation agents. Full technical specifications are available via Anthropic’s vision documentation, pending URL resolution.
/ultrareview for Claude Code CLI. Anthropic introduces the /ultrareview command for the Claude Code command-line interface, per the release announcement. Detailed documentation on command behavior is pending from Anthropic’s developer portal.
Pricing and availability, confirmed. Opus 4.7 is priced at $5.00 per million input tokens and $25.00 per million output tokens – identical to Opus 4.6. Migration carries no cost penalty for existing users. The model is available now on Claude.ai and via API across all four enterprise platforms.
| Feature | Description | Status |
|---|---|---|
| xhigh effort level | New reasoning tier between “high” and “max” | Vendor-announced; docs URL pending |
| Task budgets (beta) | Per-loop token spend caps for agentic pipelines | Vendor-announced; beta |
| Vision resolution | 2,576px max (~3.75MP), up from ~1,568px | Vendor-announced; docs URL pending |
| /ultrareview (CLI) | New command for Claude Code CLI | Vendor-announced |
| Pricing | $5/$25 per 1M tokens, parity with Opus 4.6 | Confirmed |
| Context window | 1M tokens | Confirmed |
| Platforms | API, Claude.ai, Bedrock, Vertex AI, Foundry | Confirmed |
Section 2: The Benchmark Story
This is where the deep-dive earns its length. Most AI coverage treats release benchmarks as facts. They aren’t, they’re claims at different levels of verification, and the level matters for how much you should weight them in a migration decision.
The table below maps each benchmark claim to its source, verification level, and the language you should use if you’re citing it.
| Claim | Source / Verification Level | Editorial Language to Use |
|---|---|---|
| 3x production task resolution, Rakuten-SWE-Bench | Vendor + Rakuten partner validation. Not independently confirmed. Epoch AI evaluation pending. | “Anthropic, in partnership with Rakuten, reports Opus 4.7 resolves three times more production software engineering tasks than Opus 4.6 on the Rakuten-SWE-Bench evaluation. This figure has not been independently verified by third-party evaluators at time of publication.” |
| +13% improvement, internal 93-task coding benchmark | Vendor-reported only. Benchmark methodology and sample composition not disclosed. | “Anthropic reports a 13% improvement across 93 coding tasks in its internal evaluation.” |
| 70% vs. 58%, CursorBench | Present in official release data. CursorBench operator and methodology not confirmed in available sources. | “Opus 4.7 scores 70% on CursorBench compared to 58% for Opus 4.6, per Anthropic’s release data.” |
| 98.5% accuracy, computer-use visual task assessment | Cited in release announcement. “Visual acuity” is not a standard benchmark designation. No methodology description available. | “Anthropic reports 98.5% accuracy on an internal computer-use visual task assessment.” Do not use in headlines or callouts without the methodology qualification attached. |
| Independent evaluation | Epoch AI evaluation in progress. No entry confirmed live at time of publication. | “Independent benchmarking from Epoch AI is pending. Results, when published, will be the most reliable signal for assessing whether the headline performance figures hold.” |
A note on Rakuten-SWE-Bench specifically. SWE-bench is a legitimate and widely-cited software engineering benchmark from Princeton researchers. A “Rakuten-SWE-Bench” variant appears to be a production-specific evaluation set developed in partnership with Rakuten. Partner-verified is meaningfully better than purely internal, an external organization participated in the evaluation. It isn’t the same as independent third-party confirmation. Treat the 3x figure as a credible signal worth watching, not a settled fact worth citing as proof.
CursorBench warrants similar care. Cursor is a code editor product, not an independent research body. If CursorBench is operated by the Cursor team, it sits closer to the vendor-benchmark tier than to a neutral academic evaluation. The methodology hasn’t been confirmed in available sources. The 12-point improvement over Opus 4.6 (70% vs. 58%) is noteworthy if it holds under scrutiny. Watch for methodology disclosure.
The one number that needs the most caution in any publication: 98.5% visual acuity. A specific percentage without methodology context is a credibility risk. “Visual acuity” isn’t a recognized benchmark designation. Until Anthropic publishes the evaluation methodology, treat this figure as a vendor descriptor, not a benchmark result.
Section 3: The Safety Architecture Signal
Capability announcements from frontier labs don’t happen in isolation. Anthropic’s Responsible Scaling Policy, its published framework for matching capability development to safety mechanisms, shapes how new model tiers get released. Opus 4.7’s release includes what industry coverage from Help Net Security describes as automated cybersecurity safeguards designed to limit autonomous vulnerability research. This is consistent with Anthropic’s documented approach: more autonomous capability paired with more constrained deployment conditions.
Two things to keep distinct here.
The cybersecurity safeguard framing, automated limits on what the model can do in offensive security contexts, is consistent with Anthropic’s published responsible scaling commitments and is reported by multiple industry outlets covering the release. It’s a credible characterization of Anthropic’s deployment posture, even without full technical specification from the official announcement.
The “Mythos” model class reference is a different matter. Industry coverage has referred to an unreleased model class internally described as “Mythos,” access to which may be gated by cybersecurity verification program participation. Anthropic has not publicly confirmed this characterization. It comes from T3 industry reporting only. Do not treat “Mythos” as a confirmed Anthropic product name or program.
For AI governance professionals tracking frontier lab behavior, the pattern here is worth noting regardless of the Mythos uncertainty. Anthropic appears to be building tiered access structures into its model release architecture, where demonstrated safety compliance (via a verification program) is positioned as a prerequisite for accessing more capable future tiers. That’s a meaningful governance signal whether or not “Mythos” is the real internal name.
Upstream safety research context: arXiv paper 2602.19450 has been referenced in relation to Anthropic’s safety work in this period. It predates the Opus 4.7 release and is not a technical report for this model. Include it as background reading context only, it doesn’t validate any Opus 4.7 capability claim.
Section 4: What Developers Need to Watch
1. Epoch AI independent benchmark release. This is the signal that matters most. When Epoch AI publishes its evaluation of Opus 4.7, you’ll know which headline performance figures hold and which were optimistic vendor framing. Epoch AI’s evaluations typically arrive within weeks of a flagship model release. The Rakuten-SWE-Bench 3x figure and the internal 13% coding improvement are the claims most worth watching for independent confirmation or revision.
2. Token count inflation reports. Some developers have reported potential token count increases of up to 35% for certain content types with Opus 4.7’s updated tokenizer, a claim attributed to NxCode community observations that has not been independently verified and that Anthropic has not addressed in official documentation at the time of this package. Pricing parity at the per-token level doesn’t guarantee cost parity if your actual token counts change. Enterprise users running high-volume pipelines should benchmark effective cost, not just headline rates, before committing to migration.
3. Mythos model class access pathway. Unconfirmed per industry reporting only. If the cybersecurity verification program does gate access to more powerful future model tiers, early enrollment may matter for teams that expect to need those capabilities. Watch for official Anthropic documentation before making decisions based on this.
Section 5: Migration Decision Framework
Opus 4.7 arrives approximately 70 days after Opus 4.6’s reported February 5, 2026 release. Pricing parity removes the cost barrier to testing. The practical question for teams on Opus 4.6 is whether any of the specific new features justify the migration effort for their use case. Here’s a use-case matrix based on the verified feature changes.
| Feature | Who Benefits Most | Expected Benefit | Confidence |
|---|---|---|---|
| xhigh effort level + task budgets | Teams running autonomous coding agents or multi-step agentic loops where token costs scale unpredictably | More precise reasoning control; cost-capping without degrading to “high” tier | Vendor-announced; beta feature, test before relying on in production |
| Vision resolution (2,576px) | Teams building computer-use agents, desktop automation, or UI testing pipelines | Ability to parse denser interfaces and smaller text without preprocessing crops | Vendor-announced; docs URL pending |
| /ultrareview (CLI) | Developers using Claude Code CLI for code review workflows | New review command; specifics pending documentation | Vendor-announced |
| Pricing parity | All current Opus 4.6 users | Zero cost increase to test the new version, low barrier to evaluation | Confirmed |
The migration logic is straightforward for teams that fit the top two rows: test on a representative workload before committing. The pricing parity means testing costs nothing extra. What you’re evaluating is whether the performance gains that Anthropic reports actually materialize on your specific production tasks, not on their internal 93-task benchmark.
Teams building software engineering agents on production codebases should treat Opus 4.7 as a candidate worth evaluating now. The independent benchmarking that would confirm whether it’s genuinely better than Opus 4.6 for hard production tasks is coming, probably soon. You don’t have to wait for it, but you should know it’s pending before you treat the vendor numbers as your evaluation.
The bottom line: Anthropic has shipped a model with meaningful new controls for agentic developers, a real vision resolution upgrade with concrete operational implications, and benchmark claims that are credible enough to test and not yet confirmed enough to cite. That’s a useful product release. It’s also an honest description of where the evidence stands today. Independent evaluation will sharpen the picture. Watch for it.