Anthropic released Claude Opus 5 on Thursday, July 24. Same price as Opus 4.8 — $5 per million input tokens, $25 per million output — and it's better at most things that matter for real workflows. That part is straightforward. The change worth actually understanding is the effort dial, and most teams are going to read the headline, swap the model ID, and move on without touching it.
What got better
Numbers first. On SWE-bench Multimodal — the benchmark closest to real-world coding — Opus 5 scores 59.4%. Opus 4.8 was at 38.4%. A 21-point jump on tasks that require reading visual context: fixing a bug from a screenshot of an error, not just a pasted stack trace. That shows up in agentic workflows where the model has to interpret tool output or a rendered UI before deciding what to do next.
On Frontier-Bench v0.1, the agentic terminal-coding benchmark, Opus 5 scores 43.3%. Anthropic's system card says it more than doubles Opus 4.8's score there. On BrowseComp, a web research task battery, it hits 90.8%.
The comparison to Fable 5 matters for cost planning. Fable 5 is Anthropic's frontier model and costs roughly twice per task. CursorBench 3.2 puts Opus 5 within 0.5% of Fable 5's peak. For document processing, lead qualification, content generation, or internal search — which is what most businesses are actually running — that 0.5% difference doesn't justify doubling the API bill.
Context window: 1 million tokens in, 128K max output. If your team built chunking pipelines to work around smaller context limits on older Claude models, Opus 5 is worth retesting before you keep maintaining them. You can now pass an entire contract, a year of support tickets, or a full CRM export in a single call and let the model find what it needs — no stitching required.
What the effort dial is — and why the guidance flipped
Claude has had extended thinking and token budget controls for a while now. Opus 5 formalizes this into five labeled levels: low, medium, high, xhigh, and max.
On Opus 4.x, Anthropic told API users to start at xhigh for agentic work. The model needed that headroom to handle multi-step tasks reliably. With Opus 5, that advice reversed: start at high (now the actual default), and reach down when accuracy is holding steady.
The reason this matters is cost. Effort level changes how many tokens a task consumes, even though the per-token price stays fixed. A complex reasoning task that burned 12,000 tokens at xhigh on Opus 4.8 may run just as well at 6,000–8,000 tokens at high on Opus 5. At scale, that's a real difference in the monthly bill.
The effort dial doesn't change what you pay per token. It changes how many tokens a task uses. At volume, those aren't the same number.
Which level fits which task
Low and medium are for tasks that don't require much reasoning — classification, extraction, summarization of already-structured content. An email routing workflow that sorts incoming messages into customer support, billing, and sales buckets works fine at low effort with no quality loss.
High, the new default, handles most real business work. Contract review, proposal drafting, research synthesis, lead scoring. Start here for anything where the output requires actual judgment.
Reach for xhigh or max when a wrong answer costs more than the extra compute. Legal analysis across a 200-page document. Debugging a multi-step integration failure where the root cause is ambiguous. Production agents where one hallucination triggers a support escalation. That's when you spend the tokens.
Fast mode is a separate axis — about latency, not quality. It runs Opus 5 at about 2.5x default speed for double the token price ($10 input / $50 output per million). Use it for real-time customer-facing agents where response time matters more than budget.
How to upgrade
Claude Max subscribers already have Opus 5. It's the new default and Anthropic flipped the switch automatically — no action needed.
On the API, the model ID is claude-opus-5-20260724. No tokenizer changes, no breaking API changes in this release. It's a direct drop-in for claude-opus-4-8-*. Opus 5 is also available on Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry, so existing AWS or GCP deployments don't need new infrastructure.
The practical test worth running: take your 10 highest-volume Opus 4.8 tasks, run them through Opus 5 at high effort, and compare output quality against token consumption. If quality is the same, you found a cheaper baseline. If it's better, you might be able to drop some xhigh tasks down to high. Either result is useful.
Anthropic published a detailed system card for Opus 5 alongside the launch — safety evaluations, autonomy risk testing, and the edge cases they flagged. Worth reading before you give the model write access or elevated permissions in a production agent.
For the broader model-choice question — when Haiku makes sense vs Sonnet vs Opus, and what a realistic monthly bill looks like — we covered the cost math in our AI agent cost breakdown. If you want to talk through what tier and effort level fit a specific workflow, drop us a line.
— Cole