2026-04-16 — Thursday
Anthropic released Claude Opus 4.7 today, the newest model in the Claude 4 family and now the company’s most capable generally-available model. The headline framing in the launch post is direct: Opus 4.7 is “a notable improvement on Opus 4.6 in advanced software engineering, with particular gains on the most difficult tasks.” In practice, the release is three stories stacked together — a coding-agent jump, a vision jump, and a set of new platform features for keeping long-running agents on a leash.
Anthropic positions Opus 4.7 as the model to hand off your hardest coding work to. The numbers Anthropic chose to publish on launch day are heavily weighted toward agentic-software-engineering benchmarks:
Beneath the benchmarks are two structural changes worth calling out specifically — high-resolution vision and a new tokenizer — because both affect how you actually use the model, not just how it scores.
Opus 4.7 is the first Claude model with high-resolution image support. Max image size goes from 1,568 px / 1.15 MP (Opus 4.6) to 2,576 px / 3.75 MP. Anthropic calls out two concrete use cases this unlocks:
Opus 4.7 ships with an updated tokenizer that Anthropic credits with contributing to the across-the-board quality lift. The trade-off, disclosed clearly in the announcement: the same input text maps to roughly 1.0× to 1.35× more tokens than on previous Claude models.
Practically: budget noticeably more input tokens for English prose, and significantly more for non-Latin scripts and code-heavy content. The pricing didn’t change — $5 per million input tokens, $25 per million output tokens, the same as Opus 4.6 — so the same workload may cost up to ~35% more even at identical capability. Several day-one reviewers are leading their videos with exactly this caveat (see Better Stack’s “Opus 4.7 Is GREAT (except the token usage)” in the Featured Videos below).
Two new control knobs land alongside the model:
Both controls are agent-focused. They make the most sense in long-running, autonomous setups where you can’t baby-sit individual calls and need the model to know roughly how hard it’s allowed to swing.
/ultrareview and Claude CodeFor Claude Code users specifically, Opus 4.7 ships with /ultrareview — a dedicated multi-agent review pass that flags bugs, design issues, and security smells on a branch or pull request. It’s billed separately and runs in the cloud; the local Claude Code CLI just launches it. The release post pairs this with the “Routines” feature for scheduled / 24×7 agents, which several day-one creators are pairing into “always-on coding agent” demos.
Opus 4.7 is generally available across:
API identifier: claude-opus-4-7. Pricing is unchanged from Opus 4.6: $5/M input, $25/M output. Anthropic also published a 1 M-context variant identifier claude-opus-4-7[1m] for long-window workloads — the same SKU the announcement-day Anthropic developer messaging recommends for large repos and long agentic transcripts.
The release notes describe a “similar safety profile to Opus 4.6” with low rates of deception and sycophancy, and call out specific improvements in honesty and resistance to prompt-injection attacks. The same notes flag that the model is “modestly weaker” on harm-reduction advice regarding controlled substances — a candid disclosure that’s worth reading in full if your application sits anywhere near medical or harm-reduction territory. Anthropic’s summary: “largely well-aligned and trustworthy, though not fully ideal.”
Two takeaways for practitioners:
First, the vision jump is the underrated story. The XBOW number (54.5% → 98.5%) and the resolution bump together change what computer-use and document-extraction agents can do reliably. If you parked an agentic-vision project six months ago because the model kept missing small UI text, today is the day to dust it off.
Second, the tokenizer change has cost-modelling implications. Same dollar-per-token, but up to 35% more tokens per workload, means real bills can rise even at flat capability. Re-run your token-cost projections against actual 4.7 traffic before signing any annual contracts based on 4.6 numbers.
Fireship — Viral 4-minute breakdown of the Opus 4.7 vision upgrade and what it unlocks for agents.
Nate Herk — Real-world build: Opus 4.7 wired up as a 24/7 trading agent the day after launch.
Nick Saraev — Release-day take on Opus 4.7’s benchmark gains and practical implications.
Nate Herk — Release-day skeptic angle on the “most powerful coding model” framing.
AI Explained — Performance frontier analysis plus context on the post-launch discourse.
Skill Leap AI — Hands-on walkthrough of what Opus 4.7 can do that earlier Claude versions couldn’t.
The AI Advantage — Full breakdown plus first-day testing results across coding and reasoning tasks.
WorldofAI — Side-by-side testing of Opus 4.7 against rival coding models on identical prompts.
Vaibhav Sisinty — Routines & Opus 4.7 mastery guide for automating multi-step workflows.
David Ondrej — Claude Code paired with Opus 4.7 as an end-to-end coding agent setup.
Zinho Automates — Claude Code’s biggest update: Opus 4.7 plus the new Routines for 24/7 agents.
Chase AI — Release-day reaction framing why the Opus 4.7 gap on coding feels uncatchable for now.
Developers Digest — 5-minute overview of every Opus 4.7 change that matters to developers.
Better Stack — Honest take: Opus 4.7 is great, except the new tokenizer pushes token usage up.
AI Revolution — Claude 4.7 release coverage plus discussion of Anthropic’s rumoured higher-tier model.
AI Master — Updated end-to-end tutorial on using Claude Opus 4.7 across the Anthropic stack.