Anthropic released Claude Opus 5 on July 23, 2026. Axios coverage of the release confirmed the key details: a 1 million token context window, state-of-the-art scores on coding and knowledge-work benchmarks, and a per-turn effort toggle that lets users configure how hard the model works on a given request across five levels — low, medium, high, xhigh, and max.
The benchmark scores and the context expansion are meaningful progress in a line of releases. The effort toggle is the thing that actually changes the enterprise decision calculus.
quadrantChart title Task Type vs Effort Level Selection x-axis Low Reasoning Required --> High Reasoning Required y-axis High Cost Sensitivity --> Low Cost Sensitivity quadrant-1 Opus 5 xhigh or max quadrant-2 Opus 5 medium or high quadrant-3 Haiku or Sonnet quadrant-4 Evaluate trade-off carefully Document classification: [0.2, 0.85] Contract analysis: [0.82, 0.2] Customer support routing: [0.3, 0.7] Due diligence review: [0.88, 0.25] Code generation: [0.75, 0.4] Batch summarization: [0.25, 0.75]
The rundown: what shipped on July 23
Four things worth tracking for enterprise teams:
Per-turn effort controls. The effort parameter — low, medium, high, xhigh, max — configures how much reasoning the model applies per request and, by extension, how many tokens it uses. This is not a global model setting; it can be configured per API call. Anthropic recommends starting at xhigh for agentic coding work. Set it deliberately — library defaults will not optimize this for you.
1 million token context. The model can now process an entire codebase, a full contract set, or a research corpus in a single call. For document-intensive workflows that previously required chunking and multi-call stitching, this changes what is practical.
Documented performance gains on real workloads. Box’s CTO reported 8% overall improvement over Opus 4.8, with 11% on data analysis workflows and 17% on due-diligence work. A legal AI team found Opus 5 produced 26% fewer tokens than Opus 4.8 on equivalent tasks while maintaining quality. Better output at lower token cost is the shape of a real improvement, not just a benchmark score.
Positioned below Fable 5 in cost, comparable on knowledge work. Anthropic’s framing places Opus 5 as the go-to model for most knowledge work and automation, including tasks that previously justified the Fable 5 tier. That repositions which workloads actually need Fable 5.
For the working software engineer
The effort parameter is new infrastructure to manage. It does not configure itself.
Define effort levels at the routing layer — before the API call — based on task type, not per individual request. Document classification, summarization, and extraction tasks often produce equivalent outputs at medium or high effort compared to xhigh, at measurably lower token cost. The difference is task-dependent and needs to be measured on your specific workload, not assumed from Anthropic’s benchmarks or Box’s numbers.
Anthropic’s documentation recommends starting max_tokens at 64k for xhigh and max effort settings to give the model adequate room for extended thinking and tool calls. That compounds quickly at volume. Build telemetry that reports effort level alongside token count and output quality metrics before the first production deployment — otherwise you will be diagnosing cost surprises without the data to explain them.
For business owners and operators
The effort toggle reframes the model selection conversation.
The question used to be: which capability tier — Haiku, Sonnet, Opus, Fable — fits this workload’s budget and capability requirements? Now the question is: which tier at which effort level for which task type?
This is more nuanced, but it is also more useful. A single Opus 5 deployment can serve high-volume routine tasks at low effort and high-stakes analytical tasks at xhigh on the same infrastructure. The cost profile becomes a function of workload composition rather than a fixed tier selection.
The risk is that nuance requires ownership. Teams that default to xhigh everywhere will overspend. Teams under cost pressure that constrain effort without measuring quality impact will get silent degradation — outputs that look reasonable but miss things a higher-effort call would have caught. Neither failure mode is visible without deliberate instrumentation.
My take
The more configurable AI infrastructure gets, the more the governance gap becomes the primary risk.
At LERETA, where I led architecture across a multi-year modernization of a mainframe system that processed roughly $18 billion in property taxes annually, the most valuable early work was not the technical architecture itself. It was creating a board-level picture of which systems deserved investment. Not all 30-plus applications on that platform were equal. Some were on the critical path to modernization. Others were candidates for consolidation or decommission. Making that visible — at the level where the $20 million budget decision actually got made — was what turned a vague IT initiative into a funded program.
The effort dial in Opus 5 creates the same forcing function at the infrastructure level. Every enterprise AI deployment now has a dial, and somebody has to own where it is set for each workload. If that decision is made by default rather than by policy, costs will drift, quality will be inconsistent, and no one will have a clear answer when the CFO asks why the AI spend keeps climbing.
The organizations that get this right will treat effort-level governance as a first-class infrastructure decision: documented, owned, reviewed on a cadence, and tied to both quality metrics and cost targets. Anthropic made the tool more capable and more configurable. The enterprise decision that matters is what to do with that configurability.
