Claude Opus 4.7 Launch: Anthropic Quietly Shipped the Most Capable Coding Model on the Market
What Just Happened
Anthropic released Claude Opus 4.7 today. It is generally available across Claude.ai, the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.
Pricing stays flat at 5 dollars per million input tokens and 25 dollars per million output tokens, same as Opus 4.6.
The model ID for developers is claude-opus-4-7.
The One-Line Version
Opus 4.7 is the new state-of-the-art coding model, with sharper instruction following, stronger vision, and long-horizon agent reliability that finally feels close to what the marketing has been promising for two years.
Why This Launch Is Different
Most model launches are incremental. This one is also incremental, but the places where it improves are the ones that decide whether AI agents work in production or not.
Three things stand out.
1. Instruction Following Got Strict
This is the detail people will feel first. Opus 4.7 follows instructions literally. Anthropic is openly telling developers that prompts written for older models may now produce unexpected results because the model stops interpreting loosely and starts doing exactly what you asked.
That is a big shift. Most of the implicit tricks people used to paper over weak instruction following (repeating rules, double-emphasizing, adding fallback language) now create noise instead of reliability. Teams that built prompt libraries over the last year will need to re-tune them.
2. Vision Up to 3.75 Megapixels
Opus 4.7 can process images up to 2,576 pixels on the long edge, roughly three times the resolution ceiling of previous Claude models.
This matters for a specific class of work. Computer-use agents reading dense screenshots, dashboards, and diagrams no longer have to squint. XBOW, which builds autonomous penetration testing, reported jumping from 54.5 percent to 98.5 percent on their visual acuity benchmark. That is not a small tweak. That unlocks entirely new use cases.
3. Long-Horizon Agents Actually Finish
Several partners quoted in the announcement made the same point in different words. Devin reports Opus 4.7 working coherently for hours. Notion reports a 14 percent lift with a third of the tool errors. Genspark highlights loop resistance, which is the unglamorous metric that decides whether an agent burns compute indefinitely on roughly 1 in 18 queries or actually ships.
The shift from working with agents one-to-one to managing them in parallel is a real workflow change. Opus 4.7 is aimed squarely at that shift.
What Also Launched Today
Anthropic bundled several ecosystem updates with the model itself.
A new xhigh effort level sits between high and max, giving developers finer control over the reasoning versus latency tradeoff.
Task budgets are in public beta on the API, letting developers cap Claude's token spend across long runs instead of watching runaway agents eat credits.
Claude Code got an /ultrareview slash command that runs a dedicated review pass on code changes, flagging bugs and design issues a careful human reviewer would catch. Pro and Max users get three free ultrareviews to try it.
Auto mode extended to Max users, which lets Claude Code make decisions on your behalf without stopping for permissions on every step.
The Competitive Position
Anthropic is not hiding the fact that they have a more capable model internally. Mythos Preview, announced under Project Glasswing last week, outperforms Opus 4.7 across a range of benchmarks. Mythos stays limited while Anthropic tests cyber safeguards on less capable models first.
That positioning is interesting. Opus 4.7 is effectively the highest-capability model Anthropic is comfortable releasing broadly, while Mythos sits in a controlled release because its cyber capabilities cross a threshold that requires more infrastructure.
Read between the lines and it says two things. First, the frontier is moving faster than the public release schedule. Second, Anthropic is betting that a staggered release strategy, with real safeguards tested in production on Opus 4.7 before a broader Mythos rollout, is the right approach to scaling capability responsibly.
The Migration Gotcha
One detail that will bite teams who do not read the migration guide.
Opus 4.7 uses an updated tokenizer. The same input text maps to roughly 1.0 to 1.35 times more tokens depending on content type. Combined with more thinking at higher effort levels, this means your token bills can shift even though the per-token price is unchanged.
Anthropic says net effect is favorable in their own testing, with better performance per token at all effort levels. But if you are running production workloads, measure it on your own traffic before assuming cost parity.
What It Means for Builders, Agencies, and Freelancers
For Builders Shipping AI-Powered Products
If you are using Opus 4.6 in production, this is an upgrade with no pricing change and meaningful quality gains, especially for longer tasks and multimodal work. The migration is real work if you have tuned prompts heavily, but the payoff is agent workflows that actually complete.
For Agencies and Consultancies
The instruction-following shift is the biggest talking point for client work. If you are selling AI automation or agent builds to clients, Opus 4.7 reduces the class of bugs where agents drift from the spec. Your billable hours on prompt debugging drop, and reliability goes up.
For Solo Developers and Freelancers
Claude Code's default effort level is now xhigh across all plans, and the ultrareview command is the kind of feature that replaces a senior engineering review pass for smaller projects. If you ship code alone, this is a real productivity lift.
For Business Users
GDPval-AA, which is a third-party benchmark for economically valuable knowledge work across finance, legal, and similar domains, shows Opus 4.7 as state-of-the-art. Harvey reports 90.9 percent on BigLaw Bench at high effort. Databricks reports 21 percent fewer errors than Opus 4.6 on document reasoning. These are the use cases where AI stops being a novelty and starts replacing entire workflows.
The Takeaway
Opus 4.7 is not a flashy launch. There is no new interface, no surprise product pivot, no marketing spectacle. What it does is close the gap between demo-quality AI and production-quality AI in the places that actually matter: instructions, vision, and long-horizon reliability.
The fact that Anthropic has a more capable model in the wings, gated behind safety work, says more about where the frontier is than anything in this specific release. Opus 4.7 is the model people will actually use. Mythos is a signal about what is coming next.
For anyone building on top of Claude, today is a good day to re-audit your prompts, check your token usage on real workloads, and start planning which parts of your stack can move from human-supervised to agent-completed.
The ceiling just moved up. Quietly.