The Launch
On April 16, 2026, Anthropic formally released its new flagship, Claude Opus 4.7 — just over two months after Opus 4.6. Anthropic defines it as the strongest Claude model broadly available, with across-the-board advances in coding, visual understanding, and long-task execution. Notably, Opus 4.7 is not Anthropic’s strongest hand: the internal Claude Mythos Preview posts higher numbers on every benchmark, but the company judged it “too powerful” for a public release. Opus 4.7 is therefore the strongest version users can actually touch.
The Numbers
| Benchmark | Opus 4.6 | Opus 4.7 | Gain | GPT-5.4 | Gemini 3.1 Pro |
|---|---|---|---|---|---|
| SWE-bench Verified | 80.8% | 87.6% | +6.8 | — | 80.6% |
| SWE-bench Pro | 53.4% | 64.3% | +10.9 | 57.7% | 54.2% |
| Terminal-Bench | 58.0% | 74.0% | +16.0 | — | — |
| CharXiv (visual reasoning) | 69.1% | 82.1% | +13.0 | — | — |
On SWE-bench Verified, the authoritative coding test, Opus 4.7’s 87.6 percent sets a public-model record, clearing Gemini 3.1 Pro’s 80.6 percent; on the harder SWE-bench Pro, its 64.3 percent leads GPT-5.4’s 57.7 percent.
Four Core Upgrades
1. Advanced software engineering. The centerpiece is senior-engineer behavior on long tasks: more rigorous handling of long-running work, more precise instruction following, a self-verification mechanism that checks its own output before reporting, and the ability to finish the hardest coding tasks independently. Anthropic’s user feedback says developers now confidently delegate the work that previously demanded close supervision.
2. A visual leap. Image input resolution triples — up to 2,576 pixels on the long edge, roughly 3.75 megapixels — and CharXiv visual reasoning jumps 13 points to 82.1 percent. The practical effect: drop in a dense UI screenshot and Opus 4.7 reads it pixel by pixel — small text, buttons, tables — instead of compressing away the details. Generated interfaces, slides, and documents get visibly better.
3. Agent capability evolves across the board. Multi-tool coordination improves, longer task chains hold context coherently, auto mode opens fully to Max subscribers with autonomous decision-making, and agentic financial analysis posts state-of-the-art on third-party evaluations. On the agent capability test, Opus 4.7’s max-effort tier scores 74 percent against Opus 4.6’s 61 percent — an entire tier.
4. Finer effort control. A new xhigh tier slots between high and max, letting users trade reasoning depth against latency more precisely. Claude Code’s default effort level moves up to xhigh across all plans. The official effort curve: roughly 65 at high, 71 at xhigh, 74 at max.
Safety: The Deliberate De-Scoping
Unlike Mythos Preview, Opus 4.7 intentionally reduces cyber capabilities: it ships with guardrails that auto-detect and block high-risk cybersecurity requests; “differentiated reduction” consciously weakens specific high-risk abilities; and verified security professionals can request access through a formal process restricted to legitimate research. Anthropic frames this as accumulating deployment experience ahead of releasing Mythos-class models at scale.
| Model | SWE-bench Verified | SWE-bench Pro | Terminal-Bench | Visual reasoning | Availability |
|---|---|---|---|---|---|
| Opus 4.7 | 87.6% | 64.3% | 74.0% | 82.1% | Public |
| Mythos Preview | 93.9% | 77.8% | 82.0% | 93.2% | Restricted partners |
Mythos Preview remains limited to select tech companies and financial institutions as part of Anthropic’s Project Glasswing cybersecurity program.
Pricing and Token Economics
List pricing holds at $5/$25 per million tokens (input/output). But real bills may rise: a new tokenizer makes the same text consume roughly 1.0 to 1.35 times more tokens, and deeper thinking at high effort tiers lengthens outputs. The combination nudges effective costs upward — though for developers who treat Claude as the tool of last resort for hard programming problems, the leverage from higher output quality easily outweighs the extra token spend at enterprise scale.
Ecosystem Integration
Opus 4.7 rolled out across the Claude platform (web, app, API), Claude Code, and third-party surfaces — Cursor, GitHub Copilot, Amazon Bedrock, Azure, and Google Cloud. Cursor shipped it promptly at a 50 percent discount, GitHub Copilot announced a staged rollout, and launch-day discussion saturated the tech internet.
Anthropic’s Acceleration
The cadence around this launch is its own story. Between February 1 and March 24 — 52 days — Anthropic updated 74 products, one every other day on average, spanning legal, finance, marketing, and security. Annualized revenue climbed from $9 billion at the end of 2025 to $30 billion; customers spending over $1 million a year passed 1,000, doubling in under two months; and valuation proposals jumped from $380 billion in February to roughly $800 billion. Foreign press reports an IPO possibly as soon as October, as the listing race with OpenAI turns white-hot.
The Controversies
The release is not without criticism. The “model got dumber” question: AMD’s AI team, analyzing 6,800 Claude Code session files, found that “thinking shadowing” correlates strongly with quality regressions on complex long-session engineering tasks, that thinking length dropped 73 percent versus earlier versions, and that the model’s internal thinking depth itself may be declining — not just its transparency to users. The tokenizer’s context cost: veteran developer Simon Willison noted the same text now splits into up to 1.35 times more tokens, so developers hit context limits sooner and prompts grow more bloated, compounding context rot. Multi-platform variance: Anthropic simultaneously serves inference on AWS Trainium, Google TPUs, and some NVIDIA GPUs — meaning different tool calls within one Claude Code session can land on different cloud platforms with different service environments, producing quality wobble across a single task flow.
The Bottom Line
Opus 4.7 is Anthropic’s careful balance of performance and responsibility: record-setting coding, tripled visual acuity, an agent tier that jumped 13 points, and finer effort control — set against deliberately de-scoped cyber abilities, a 1.0-1.35x token inflation, and unresolved questions about thinking depth. It fits developers with complex programming tasks, users who need high-fidelity visual understanding, teams running long autonomous agent workflows, and enterprises that weigh model safety heavily. The deeper signal: with Mythos Preview held back, the AI race has moved past parameters and fundraising into revenue, governance, and IPO timetables — and Anthropic just moved all three.
Learn more: Anthropic’s official blog. Opus 4.7 is now served through the Claude API, Amazon Bedrock, Azure, and Google Cloud.
