Claude Opus 4.7: Pricing, Benchmarks & Mythos Comparison

A same-price upgrade, not a new tier

Anthropic made Claude Opus 4.7 generally available on April 16, 2026, through Claude, the API, and Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry. The API identifier is claude-opus-4-7. [1]

Standard API pricing held at $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.6. Anthropic shipped the release as an incremental update to an existing model line rather than as a new pricing tier. [1]

New ways to spend a request's budget, not just its tokens

The release added an xhigh effort setting between the existing high and max levels, giving API callers a finer step between a fast answer and a maximally thorough one. It also opened task budgets in public beta, letting a developer cap how many tokens a request may spend rather than only choosing an effort level. [1]

In Claude Code, Anthropic added an /ultrareview command for a dedicated review pass and extended auto mode, previously limited to other plans, to Max plan subscribers. Vision input rose to images up to about 2,576 pixels on the long edge, which Anthropic described as more than three times the resolution prior Claude models accepted. [1]

Atlas interpretation: Effort levels and task budgets answer different questions. Effort sets how hard the model tries; a budget sets how much that effort is allowed to cost. Splitting them lets a caller ask for maximum thoroughness on a request while still capping what a runaway chain of tool calls can spend. [1]

The gains are measured against the model it replaces, on partner-run tests

Anthropic reported 87.6 percent on SWE-bench Verified, with a footnote noting that problems flagged by its memorization screens were excluded and that Opus 4.7's margin over Opus 4.6 held after that exclusion. On CursorBench, Opus 4.7 reached 70 percent against 58 percent for Opus 4.6. On Terminal Bench 2.0, Anthropic reported a 13 percent improvement over Opus 4.6, including four tasks the earlier model could not solve at all. [1]

Several of the other cited scores came from enterprise partners testing the model on their own workloads: Harvey's BigLaw Bench at 90.9 percent at high effort, Rakuten reporting three times as many production coding tasks resolved as Opus 4.6, and Databricks reporting 21 percent fewer errors on a document question answering benchmark. [1]

Atlas interpretation: A partner's own benchmark is evidence about that partner's workload, not an independent audit. Harvey, Rakuten and Databricks each have a commercial relationship with Anthropic and picked the task their own product needed to measure, so a reader should treat these as vendor-selected comparisons against Anthropic's prior model rather than as neutral third-party scores. [1]

Anthropic named the model it is not selling

Anthropic stated in the same announcement that Opus 4.7 is less broadly capable than its most powerful model, Claude Mythos Preview, and specifically that Opus 4.7's cyber capabilities are not as advanced as Mythos Preview's. It also stated that Mythos Preview remains the best aligned model it has trained, by its own evaluations. Anthropic had restricted Mythos Preview to an invitation-only program for critical infrastructure operators nine days earlier, in Project Glasswing, rather than releasing it generally. [1]

Atlas interpretation: The comparison is voluntary. Anthropic did not have to name Mythos Preview in a launch announcement for a different model, and doing so put a ceiling on Opus 4.7's own release on the day it shipped. Framed against a model nobody outside the vetted program can buy, the admission reads less like modesty and more like a way to advertise Mythos Preview's capability without lifting the restriction on it. [1]