Claude Opus 5 is Anthropic's new flagship in the Opus line, released on July 24, 2026, succeeding Claude Opus 4.8. Anthropic positions it "for complex agentic coding and enterprise work," and it is now the default model in Claude Code, Anthropic's own coding agent. It is available through the Claude API (model ID claude-opus-5), Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, and Claude Platform on AWS.
What's new
Anthropic's own framing is that Opus 5 "matches many of Fable 5's capabilities at half the price" — a claim worth noting as Anthropic's characterization rather than an independently verified benchmark, since no third-party comparison numbers accompany it.
Under the hood, Opus 5 replaces the older extended-thinking mechanism — where developers set a manual budget_tokens value — with adaptive thinking. There is no manual thinking-budget parameter in Opus 5. Instead, the model exposes an effort parameter with five settings: low, medium, high, xhigh, and max, defaulting to high. This shifts the control surface from "how many tokens should the model spend thinking" to "how much effort should the model apply," letting Anthropic tune the thinking-to-output ratio internally per effort tier rather than exposing a raw token budget.
Training data coverage extends through May 2026, which Anthropic states is the most recent cutoff of any Claude model currently shipping. For anyone using the model for research, current-events reasoning, or dependency/library awareness, that puts Opus 5 closer to the present than any prior Claude release.
Long-horizon autonomy
Anthropic's headline claim for Opus 5 is durability on long, unsupervised tasks. According to Anthropic, the model can work independently for hours on a single task, recover from its own errors mid-task, route around blockers instead of stalling out, and periodically check its own work rather than pushing forward on a possibly-wrong path without verification.
Anthropic backs this with two benchmark results it reports itself: 43.3% on Frontier-Bench v0.1, up from Opus 4.8's 18.7%, and 68.8% on DeepSWE v1.1, up from Opus 4.8's 59.0%. Both of these are Anthropic's own published figures, not figures independently reproduced by Tokonomix's test infrastructure — they should be read as the vendor's account of its own model's progress on its own agentic-coding benchmarks, not as a neutral third-party result. Anthropic frames Frontier-Bench specifically as a measure of exactly this cluster of behaviors: sustained task execution, self-correction, and autonomous recovery, rather than raw single-turn accuracy.
The size of the jump matters more than the absolute numbers here. More than doubling a score on a benchmark built around sustained, multi-step execution (18.7% to 43.3%) is a different kind of claim than a few points of accuracy gain on a static question-answering set — it is specifically a claim about task duration and error-recovery, not about single-turn reasoning quality. The DeepSWE v1.1 gain (59.0% to 68.8%) is a smaller relative jump but sits on an already-high base, which is consistent with Anthropic's narrative that this release targets the tail end of agentic coding — the point where a model has to keep going after something has already gone wrong, rather than the front end of getting a first attempt mostly right.
Judgment and clarifying questions
The second capability Anthropic highlights is improved judgment before acting. Per Anthropic, Opus 5 is more inclined to ask a clarifying question before guessing at an ambiguous instruction, more willing to push back when given a flawed or underspecified instruction rather than executing it literally, and more likely to reason through the downstream implications of a change before implementing it.
For agentic coding specifically — the use case Anthropic frames Opus 5 around — this matters because a model that silently guesses at ambiguous requirements can burn a long autonomous session executing the wrong plan. A model that asks first, or flags an assumption before acting on it, fails cheaper. Anthropic does not publish a quantitative benchmark for this behavior; it is described qualitatively as a design intent for the release rather than measured with a numeric score.
Read together, the long-horizon autonomy claim and the judgment claim describe two different failure modes that Anthropic says Opus 5 addresses. The first is about what happens after a mistake has already been made mid-task — does the model notice, recover, and keep going, or does it stall or compound the error. The second is about preventing avoidable mistakes before they happen — does the model ask when it should, or does it plow ahead on an assumption. Anthropic positions both as central to why Opus 5 is built for longer, less-supervised coding sessions rather than short single-turn exchanges.
Context window and pricing
Opus 5 carries a 1,000,000-token context window, with a maximum synchronous output of 128,000 tokens per response. In practical terms, a 1M-token window is large enough to hold sizeable codebases, long document sets, or extended multi-file review sessions in a single context without external retrieval — a relevant capability for the "complex agentic coding" positioning, where a coding agent frequently needs to reason over many files at once rather than one snippet at a time. The 128k output ceiling caps how much a single synchronous response can return, which matters for workflows that expect a large generated artifact — e.g., a big refactor diff or a long generated document — in one call.
Pricing is identical to Opus 4.8, so upgrading to Opus 5 does not change the cost side of an existing integration — only the model's behavior and knowledge cutoff change.
When to pick Opus 5 vs Fable 5 vs Sonnet
Anthropic's own positioning frames Opus 5 as the choice for complex agentic coding and enterprise work, and as the new default inside Claude Code — suggesting Anthropic sees it as the right general-purpose choice for coding-agent workloads going forward.
For workloads that do not need frontier-level reasoning or long unsupervised autonomy — simpler classification, shorter-turn chat, lighter-weight extraction — a smaller, lower-cost tier in Anthropic's lineup will typically be the more cost-efficient choice; Opus 5's pricing and capability profile are built around sustained, complex, higher-stakes work rather than high-volume simple tasks.
A practical way to frame the decision: if the workload is a coding agent that needs to run for an extended stretch with minimal supervision — multi-step refactors, long debugging sessions, or repository-scale changes where the model has to keep working after hitting an unexpected obstacle — Opus 5's positioning and its stated autonomy gains over Opus 4.8 are directly aimed at that case. If the workload is bounded, single-turn, or does not depend on the model recovering from its own mistakes over a long session, the price difference between Opus 5 and a lighter Anthropic model is unlikely to be justified by capability actually used.
Availability on Tokonomix
Claude Opus 5 is listed on Tokonomix under the slug claude-opus-5, tier A, hosted in the US via Anthropic direct — consistent with Opus 5's broader availability through the Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, and Claude Platform on AWS. Independent benchmark scores from Tokonomix's own test runners are not yet available for this model — they will appear automatically on this page once Claude Opus 5 is included in the next scheduled test cycle, alongside the same head-to-head comparisons already published for other catalog models.