Claude Opus 4.7 Is Generally Available — And It’s Not a Minor Bump
Anthropic shipped Claude Opus 4.7 on April 16. Benchmarks land where the leaks suggested: SWE-bench Verified climbs from 80.8% to 87.6%, SWE-bench Pro from 53.4% to 64.3%, CursorBench from 58% to 70%. GPQA Diamond comes in at 94.2%, statistically tied with Gemini 3.1 Pro (94.3%) and GPT-5.4 Pro (94.4%). The real surprise is vision: acuity jumps from 54.5% to 98.5% on XBOW’s benchmark, input resolution triples to ~3.75 megapixels, and CharXiv chart-reasoning moves 13 points without tools. Pricing holds at $5/$25 per million tokens.
New this release: an xhigh effort level that slots between high and max (Claude Code now defaults to xhigh on all plans), plus a new tokenizer that can produce 1.0x–1.35x more tokens for the same input text. Opus 4.7 is live in Claude Code, Claude apps, GitHub Copilot, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry on day one. Anthropic is recommending prompt retuning: 4.7 follows instructions more literally than any previous Claude, so sloppy phrasing that 4.6 interpreted charitably now gets taken at face value.
Anthropic Concedes Opus 4.7 Trails Mythos — By Design
For the first time, Anthropic is explicitly positioning a public model as a step below its strongest internal one. CNBC and Axios both reported the framing: Opus 4.7 is the most capable generally available Claude, but Mythos Preview — the cyber-capable model behind Project Glasswing — is stronger overall and shows the lowest rates of misconduct in internal evals. Anthropic says it deliberately reduced Opus 4.7’s cyber capabilities versus Mythos and layered on new classifiers that detect and block prohibited or high-risk cybersecurity requests.
This is the first clean articulation of what the two-tier model strategy looks like in practice: the frontier model stays private and consortium-only; the product model ships behind automated refusal layers for a specific risk class. The model card makes this explicit. If Opus 5 follows the same pattern, expect more capabilities to ship with policy-level guardrails wired in at the serving layer.
Claude Code Review Ships to Team and Enterprise — 54% of PRs Now Get Substantive Review
Anthropic formally launched Code Review for Claude Code alongside the Opus 4.7 drop. The setup: a fleet of specialized agents dispatch when a PR opens, hunt for bugs in parallel, verify findings to cut false positives, then rank and post a single overview comment plus inline findings at the exact lines. Each finding includes a collapsible “extended reasoning” section so reviewers can see why Claude flagged it. Internal metrics: before, 16% of Anthropic PRs received substantive review comments; now it’s 54%.
Admin setup is one-click in Claude Code settings plus the GitHub App install. Reviews run automatically on new PRs, billed on token usage at roughly $15–$25 per review depending on diff size. Available today as a research-preview beta to Team and Enterprise plans only — Pro is not in scope for this release.
Advisor Strategy, Monitor Tool, and Managed Agents: The Real Product Story
The model drop grabbed the headlines; the three agent primitives underneath it are what enterprise buyers will spend 2026 budget on. Advisor formalizes a pattern devs were already hand-rolling: a fast/cheap executor (Sonnet or Haiku) consults a higher-intelligence advisor (Opus) mid-generation only when it hits hard problems. Anthropic’s eval: Sonnet with Opus-as-advisor beats Sonnet alone by 2.7 points on SWE-bench Multilingual while cutting per-task cost by 11.9%. Monitor Tool replaces the “sleep then poll” loop with interrupt-driven callbacks when background processes finish, saving the tokens that used to go into status probes.
Managed Agents is the biggest of the three: a fully managed agent harness that runs Claude as an autonomous agent with secure sandboxing, built-in tools, and server-sent event streaming. Early customers tell Anthropic they’re shipping 10x faster. This is the long-rumored “skip-the-harness” release — devs who have been rebuilding the same execution loop over LangChain, CrewAI, and home-grown Python every quarter now have a first-party option. The question: how much lock-in comes with the sandbox.
Opus 4.7 Developer First Impressions: “Genuinely Better, Genuinely More Expensive”
Twelve hours of Hacker News and Reddit hands-on tracks in one direction: the gains are real and the token bill is bigger. DEV Community’s 6-hour stress test and Decrypt’s review both call out the same pattern — the model handles hard, long-running coding work more autonomously than 4.6 did, including tasks 4.6 and Sonnet 4.6 couldn’t solve at all. Warp’s early-access report: Opus 4.7 passed Terminal-Bench tasks that previous Claudes failed, including a concurrency bug 4.6 couldn’t crack. CodeRabbit notes actionability on flagged issues jumped from 54% to 64%.
The cost story is the other half. Between the new tokenizer (+0–35%) and xhigh default in Claude Code, developers are reporting real Opus bills running 15–40% higher for comparable workloads. Decrypt’s headline — “token-eating machine” — is harsh but lands. Finout’s pricing analysis puts output-token inflation as the dominant cost driver since outputs are priced 5x inputs. Practical advice circulating in dev channels: start with high, only escalate to xhigh when the task justifies it, and measure before you migrate.
Two Days Until Claude Haiku 3 Hard Cutoff — April 19
Two days left on the claude-3-haiku-20240307 deprecation clock. After April 19, API calls against the old model string will fail outright. Drop-in migration: Claude Haiku 4.5 (claude-haiku-4-5-20251001) — faster, more capable, same price tier. If you still have cron jobs, batch pipelines, or legacy agents pinned to Haiku 3, this is the last business week to migrate cleanly.
Also on the calendar: the 1M-token context beta for Sonnet 4.5 and Sonnet 4 expires April 30, and Opus 4 plus Sonnet 4 both sunset June 15. Migration targets are Opus 4.7 (new default) and Sonnet 4.6 — same API shape, better evals, lower effective price when paired via the Advisor strategy.
The Register: Anthropic Ejects Bundled Tokens From Enterprise Seat Deal
The Register broke it clean: Anthropic has unbundled usage from its $20/seat/month enterprise tier. Previously, a chunk of Claude tokens came included with the seat fee; on contract renewal, that allowance is gone. Every token now bills at standard API rates on top of the base. A company spokesperson confirmed the transition started quietly in November 2025, accelerated with the unified $20/seat plan in February, and is now the default for new and renewing enterprise contracts.
Customer impact is uneven. Implicator and PYMNTS note that for heavy accounts, the bundled allowance was already a small fraction of the bill — maybe 20% — so unbundling is mostly cosmetic. For lighter seat-heavy customers who were effectively overpaying for seats and underusing tokens, the math flips the other way. The broader read: this is Anthropic confident enough in its $30B run-rate to stop subsidizing per-seat economics. OpenAI’s Microsoft-bundled ChatGPT Enterprise remains flat-rate. The competitive pressure on procurement teams just shifted.
TrendAI Embeds Claude Across Security Platform in New Anthropic Partnership
TrendAI announced a strategic engagement with Anthropic this morning, embedding Claude models across its platform to power agentic security workflows, automation, and AI-native threat research. Claude will sit under TrendAI Vision One and plug into the company’s Zero Day Initiative and Pwn2Own programs — which is to say: a commercial security vendor whose core product is finding bugs is now running on the same model family that Anthropic has explicitly held back from general release for cyber-capability reasons.
Anthropic’s positioning is consistent — Opus 4.7 ships with cyber-use safeguards, Mythos stays behind Glasswing, partners on the consortium side get privileged access under contract. The TrendAI deal is the ecosystem version of that strategy: commercial use of cyber-adjacent capability requires a named enterprise relationship, not an API key. Expect more of these deal shapes as the offensive/defensive line keeps moving.
Opus 4.7 Is the Wrapper. Managed Agents Is the Product.
Zoom out on yesterday’s launch package and the benchmark graphs are a distraction. SWE-bench Verified going from 80.8 to 87.6 is real and impressive, but the competitive picture hasn’t moved much — on GPQA Diamond, Opus 4.7, GPT-5.4 Pro, and Gemini 3.1 Pro are inside 0.2 points of each other. The frontier labs are converging on raw reasoning. Where they’re differentiating is the harness. Anthropic shipped three pieces of harness yesterday — Advisor, Monitor, Managed Agents — plus Code Review, plus Routines three days ago, plus the upcoming design tool. That stack is what enterprise CTOs are actually procuring in 2026.
The unbundling story sits in the same frame. When you’re selling access to a raw model, flat per-seat pricing makes sense. When you’re selling managed agent infrastructure that runs on your cloud, uses your sandbox, consumes unpredictable amounts of compute, and delivers deterministic business outcomes, pure metered billing is the only rational shape. Anthropic isn’t raising prices; it’s re-pricing to match the product it’s actually becoming — an agentic-work utility, billed like AWS. The flat-fee era ending isn’t bad news for customers. It’s a tell about what Anthropic thinks the next 18 months look like.