Overview
Claude Opus 4.7 is Anthropic's April 16, 2026 flagship upgrade, landing on the same $5/$25 per million token pricing as Opus 4.6 but pushing agentic coding, vision, and long-horizon reliability meaningfully further. Opus 4.7 targets teams running production coding agents, parallel agent orchestration, and multimodal analysis over dense technical content — chemical structures, patent diagrams, full-page screenshots — where the earlier Opus generation hit accuracy or resolution limits.
The model is available the day of release on claude.ai, Claude Code, the Anthropic API as claude-opus-4-7, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Two behavior changes warrant attention before upgrading: the tokenizer now produces roughly 1.0–1.35× more tokens for the same input, and instruction-following is more literal — existing prompts that relied on implicit inference may need retuning.
What's New
Production-Grade Agentic Coding
Opus 4.7 deepens Anthropic's existing Opus push into long-running agentic work. Opus 4.6 had already emphasized agent teams, adaptive thinking, and longer autonomous coding tasks. Rakuten's internal SWE benchmark saw Opus 4.7 resolve roughly 3× more production tasks than Opus 4.6, and CursorBench climbed from 58% with Opus 4.6 to 70%. CodeRabbit reports more than a 10 percentage point recall improvement for bug detection on its review pipeline. The model also generates cleaner code — fewer wrapper functions, less dead scaffolding — and self-corrects more reliably when tool calls error out mid-task.
3× Higher Vision Resolution
Opus 4.7 accepts images up to 2,576 pixels on the longest edge (~3.75 megapixels), more than 3× the ceiling of previous Claude models. Paired with sharper perception, the XBOW visual acuity evaluation jumped from 54.5% on Opus 4.6 to 98.5% on Opus 4.7. This removes the previous downsampling penalty for dense inputs: fine print in contracts, full-page UI screenshots for browser agents, multi-panel chemical structures, and technical engineering diagrams are now usable without pre-cropping.
New xhigh Effort Control
A new reasoning effort level, xhigh, sits between high and max. It gives teams a finer latency/quality tradeoff for workloads where high is no longer enough but max is prohibitively expensive — for example, agentic workflows that need stronger planning on a per-step basis without paying full max latency across an entire session.
More Literal Instruction-Following
Anthropic explicitly calls out that Opus 4.7 interprets directives "more literally" than prior Claude models. The model is less likely to reinterpret or soften user instructions based on implicit context. This is a net win for agentic pipelines where instruction drift previously caused silent failures, but it means prompts that relied on Claude "reading between the lines" will need retuning (see Migration Guide below).
Multimodal & Domain Strength
Beyond raw vision resolution, Opus 4.7 shows Anthropic's highest internal scores to date on a Finance Agent evaluation covering document reasoning, data extraction, and multi-step modeling — positioning it against domain-specialized agents rather than just general chatbots. Anthropic also introduced a Cyber Verification Program alongside this release, enabling vetted security researchers to use Opus 4.7 for legitimate penetration-testing workflows.
Performance Benchmarks
Headline numbers from Anthropic's release (all vs. Opus 4.6 unless noted):
| Benchmark | Opus 4.7 | Prior Baseline | Δ |
|---|---|---|---|
| CursorBench | 70% | Opus 4.6: 58% | +12 pp |
| Rakuten-SWE-Bench (production tasks) | ~3× resolution | Opus 4.6 | 3× more tasks solved |
| CodeRabbit recall (bug detection) | +10+ pp | Prior best | 10+ pp recall improvement |
| XBOW visual acuity | 98.5% | Opus 4.6: 54.5% | +44 pp |
| Finance Agent eval | State-of-the-art | Opus 4.6 | New high |
The coding numbers are most relevant for teams running AI code checker pipelines or autonomous coding agents; the XBOW jump is the single largest per-category gain and reflects the vision resolution increase, not just reasoning improvements.
Migration Guide
Two behavior changes require attention when moving from Opus 4.6 (or earlier) to Opus 4.7:
Tokenizer Change
Opus 4.7 ships with an updated tokenizer. The same input maps to roughly 1.0–1.35× more tokens than Opus 4.6 depending on content type — code and structured data are at the higher end of that range. Practical implications:
- Re-estimate per-request input cost before cutting traffic over, especially for workloads dominated by code or JSON.
- Revisit any hard-coded context budgets, prompt splitters, or chunkers that assume Opus 4.6 token counts.
Prompt & Harness Retuning
Because instruction-following is more literal, prompts that previously relied on implicit inference may under- or over-perform. Anthropic explicitly recommends re-tuning prompts and agent harnesses when upgrading. In practice:
- Spell out assumptions that used to be implicit ("use only the tools provided", "do not modify files outside /src", etc.).
- A/B test high-volume prompts against Opus 4.6 on a sample of real traffic before full cutover.
- Review guardrails in agent loops — stricter adherence can surface previously masked edge cases.
API identifier: claude-opus-4-7. Opus 4.6 remains available for teams that need a gradual migration window.
Pricing & Plans
Opus 4.7 pricing is unchanged from Opus 4.6:
| Plan | Price |
|---|---|
| API – Input | $5 per million tokens |
| API – Output | $25 per million tokens |
| Claude.ai Pro | Included within plan limits |
| Claude.ai Max | Included with higher usage allowances |
| Team / Enterprise | Contact Anthropic for volume pricing |
| Cloud availability | Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry (same provider pricing as Opus 4.6) |
Effective cost note: Because the new tokenizer produces 1.0–1.35× more tokens for the same input, per-request spend may rise slightly even though the per-token price is flat. Budget with the higher end of that range for code-heavy workloads.
Best For
- Engineering teams running autonomous coding agents or multi-agent orchestration in production, where Opus 4.6 hit reliability or code-quality ceilings
- Developers using Claude Code on large codebases who want fewer wrapper functions and stronger self-correction
- Finance, legal, and research teams analyzing dense visual documents — chemical structures, multi-panel diagrams, full-page screenshots — where the 3× vision resolution matters
- AI agent platforms that need tighter instruction-following and fewer silent prompt interpretations across long tool-use loops
- Teams already evaluating Claude Sonnet 4.6 for cost reasons but needing Opus-tier depth on the hardest 10–20% of their workload
- Security research teams accepted into the Cyber Verification Program for sanctioned penetration-testing work
FAQ
How is Claude Opus 4.7 different from Opus 4.6?
Opus 4.7 delivers its largest gains in agentic coding (CursorBench 58% → 70%, Rakuten-SWE 3× task resolution), vision (XBOW 54.5% → 98.5%, 3× image resolution ceiling), and reasoning effort control (new xhigh level). Pricing is unchanged at $5/$25 per million tokens. The main trade-off is a tokenizer change (~1.0–1.35× more tokens per input) and more literal instruction-following that may require prompt retuning.
Do I need to change my existing prompts when upgrading?
Anthropic explicitly recommends re-tuning prompts and agent harnesses. Opus 4.7 interprets instructions more literally, so prompts that relied on implicit context may need to state assumptions explicitly. A/B testing high-volume prompts against Opus 4.6 on a sample of real traffic before full cutover is the safest approach.
Will my API costs go up if the per-token price is the same?
Effective per-request cost can rise by roughly 0–35% because the same input produces more tokens under the new tokenizer. Code-heavy and structured-data workloads sit at the higher end of that range. Budget accordingly before shifting large traffic volumes.
Where can I use Claude Opus 4.7?
Available on Claude.ai (Pro, Max, Team, Enterprise plans), Claude Code, and the Anthropic API as claude-opus-4-7. Amazon Bedrock and Google Cloud Vertex AI both listed Opus 4.7 as available on the April 16, 2026 release day. Microsoft Foundry currently labels Opus 4.7 as preview, so availability on that partner cloud may differ from the others at launch.
When should I still use Opus 4.6 or Sonnet 4.6 instead?
Stick with Opus 4.6 if you cannot absorb the prompt retuning window or the tokenizer cost shift on a short timeline. Choose Sonnet 4.6 when you need near-Opus coding quality at $3/$15 per million tokens and do not require the vision resolution or deepest reasoning that Opus 4.7 offers.
Is Opus 4.7 safe for sensitive security work?
Anthropic introduced a Cyber Verification Program alongside Opus 4.7 to enable vetted security researchers to use the model for legitimate penetration-testing workflows. Unverified accounts are subject to Anthropic's standard usage policies. See Anthropic's system card and the Cyber Verification Program documentation for full eligibility and safety details.
