Anthropic released Claude Opus 4.7 on April 16, 2026. Pricing is unchanged from Opus 4.6 — $5 per million input tokens, $25 per million output — and so is the context window at 1M tokens. What changes is performance across vision, coding, and document reasoning, along with a set of agentic capabilities and one significant infrastructure change that every enterprise team needs to test before migrating.
Where the performance gains land
The headline improvement is vision. Opus 4.7 scores 98.5% on the visual acuity benchmark, up from 54.5% for Opus 4.6 — a 44-point jump. The model now supports images up to 3.75 megapixels (2,576 pixels on the longest edge), three times the previous limit. For agents that process screenshots, invoices, technical diagrams, or scanned documents, this is a step change rather than an incremental improvement. Tasks that previously required pre-processing to upscale or segment images can now pass source material directly.
Coding performance improved 13% on a 93-task benchmark, but the more meaningful signal for enterprise agentic work is the Rakuten-SWE-Bench result: Opus 4.7 resolves three times more tasks than its predecessor on that production-grade coding workload. The benchmark measures multi-step software engineering tasks drawn from real repositories — closer to what autonomous coding agents actually encounter than sanitised benchmark suites. Stripe, CodeRabbit, and Cursor are among the early adopters deploying Opus 4.7 for software engineering work.
Document reasoning improved by 21% on OfficeQA Pro, a benchmark of dense multi-document reasoning over the kind of content enterprises actually work with: financial reports, legal filings, technical specifications. Some of that improvement reflects the extended knowledge cutoff — January 2026 versus May 2025 for Opus 4.6 — and some reflects stronger reasoning. Anthropic also reports state-of-the-art results on a Finance Agent evaluation; Harvey is an early adopter for legal knowledge work, and Devin and Notion are building agent capabilities on the model.
The agentic additions
Three capability changes matter specifically for enterprises building agentic systems.
The new xhigh effort level sits between the existing high and max settings, giving you a finer-grained tradeoff between reasoning depth and latency. Claude Code has raised its default effort level to xhigh across all plans. For enterprise workflows where max produces unnecessary latency and high isn’t thorough enough, xhigh is likely the right production default for complex reasoning tasks. The max level remains available for the slowest, most critical decisions where token spend is acceptable.
Task budgets are now in public beta. The feature lets you specify how Claude should allocate its thinking tokens across a long-horizon task — which sub-steps warrant deep reasoning and which can be handled quickly. This is relevant for multi-step orchestration where token spend is difficult to predict without explicit guidance. Cost predictability in long-running agentic workflows is one of the harder enterprise requirements to satisfy; task budgets are the first native mechanism in the API that directly addresses it.
Memory handling across sessions has improved. Opus 4.7 makes better use of file system-based context, retaining notes and task state more effectively across multiple sessions. Combined with the unchanged 1M token context window, this improves coherence in workflows that span more than a single conversation — exactly the long-horizon autonomous work that Anthropic specifically highlights as a target use case for the model.
The tokenizer change
This is the item that requires the most operational attention before migration.
Opus 4.7 ships with a new tokenizer. The same input text maps to 1.0–1.35× more tokens depending on content type. Per-token pricing is unchanged, but actual costs on real traffic will increase. The effect is larger for structured content — code, JSON, YAML, XML — than for plain prose. A prompt that was optimised to stay under a cost threshold on Opus 4.6 may exceed it on Opus 4.7 without any change to its content.
Teams should test token consumption on representative production traffic before cutover, not on a synthetic benchmark. Monitor per-request costs for the first two weeks after migration. If you are running agents with tight token budgets or cost-sensitive orchestration loops, factor in a worst-case 35% increase in input costs when modelling migration economics.
Opus 4.7 also follows instructions more literally than its predecessor. That is generally beneficial — more predictable behaviour is what most enterprise deployments want — but prompts that relied on the model inferring intent from loose language may produce different outputs. Test your most important prompt templates against the new model before switching production traffic.
What doesn’t change
Context window stays at 1M tokens, with no long-context premium. Pricing is identical to Opus 4.6. The full feature set carries over: prompt caching, batch processing, Files API, PDF support, computer use, MCP connector, and memory. The maximum synchronous output is unchanged at 128k tokens; the 300k output limit via the output-300k-2026-03-24 beta header is now available on Opus 4.7 (it was limited to 32k on Opus 4.6).
Opus 4.7 supports Adaptive Thinking but not Extended Thinking — that distinction is unchanged from the previous generation. Extended Thinking remains available on Sonnet and Haiku models.
Agent identity and authorization infrastructure you have built for Opus 4.6 carries forward without modification. The model ID changes to claude-opus-4-7; everything else in how you connect, authenticate, and govern the model is the same.
On Opus 4.0 deprecation
If your deployment still runs Claude Opus 4.0, June 15, 2026 is the hard deadline. Opus 4.0 had a 200k token context window; Opus 4.7 gives you 1M tokens — five times larger — at meaningfully better performance across every evaluated dimension. The migration path is straightforward but requires testing given the instruction-following and tokenizer changes described above. Don’t leave it until June.
The broader signal
The pattern across this release is consistent with where Anthropic has been heading: higher performance on the tasks that define production agentic value — long-horizon coding, dense document reasoning, multi-session memory — paired with tighter control primitives for enterprise cost management (task budgets, the xhigh effort level). The 1M context window at standard pricing, introduced in Opus 4.6 and unchanged here, remains one of the most operationally significant differences between Claude and comparable frontier models for the kind of large-context enterprise workflows that benefit most from a wide context.
The vision improvement deserves attention even from teams that don’t currently process images. At 98.5% visual acuity accuracy — nearly double the previous generation — the case for including screenshots and diagram images as native inputs to agentic workflows rather than pre-processing them into text is materially stronger. If you shelved image-in-context workflows because accuracy wasn’t there, this is the release to revisit that decision.