Claude Opus 5 changes the Claude default

Claude Opus 5 is the new default Claude model for most coding and agent workloads.

Anthropic released it on July 24, 2026 at $5 per million input tokens and $25 per million output tokens—the same base price as Opus 4.8.

Its 96.0% SWE-bench Verified result leads Fable 5, but Fable still wins the harder SWE-bench Pro row. The upgrade case is strong, not universal.

The same price makes Opus 4.8 hard to defend

Anthropic did not charge a generation tax. Opus 5 keeps the claude-opus-5 API model at the same $5/$25 rate as Opus 4.8, while Fast mode runs at roughly 2.5 times the default speed for twice the token price. Opus 5 ships across the Claude API, Claude Code, Claude Pro, and Claude Max. Max uses it as the default.

The API model ID is claude-opus-5, with a 1 million-token context window. Anthropic exposes low, medium, high, xhigh, and max reasoning effort. Thinking cannot be disabled at xhigh or max, so those settings change more than a slider label.

Current rates live on the Claude pricing owner and in the cost calculator. Anthropic's official launch post is the source for availability, effort settings, and Fast mode.

Price changes the migration burden. Teams moving from Opus 4.8 can test a stronger successor without first renegotiating the per-token budget. Compatibility remains the honest limit: a benchmark gain does not prove that one production prompt, tool schema, or approval loop will behave the same. Run the regression set before replacing the model ID.

Token price is not task price. Max effort may use more reasoning tokens, Fast mode doubles the unit rate, and retries can erase a cheap headline. A useful migration test records input tokens, output tokens, turns, retries, and wall time beside task success. Fewer successful turns can beat a cheaper row. A looping agent cannot.

The headline win is narrower than it looks

We checked the launch claims against the Claude Opus 5 system card, the verification registry, and the current model row. Six results explain most of the release without pretending every percentage belongs on one scale.

Table 1 Download CSV

Evaluation	Opus 5	What the row can support
SWE-bench Verified	96.0%	Strong result on the established verified issue-resolution set
SWE-bench Pro	79.2%	Harder repository work; Fable 5 remains slightly ahead
FrontierCode 1.1 Main	53.4%	Near-parity with Fable 5 on a separate coding protocol
FrontierBench v0.1	43.3%	Anthropic internal run, not a benchmark-native leaderboard row
OSWorld 2.0	70.6%	Computer-use result on the 2.0 protocol
ARC-AGI-3	30.16%	Verified high-effort result; no max-effort row was available at launch

That 96.0% result deserves the headline. Anthropic averaged five trials, and the source maps cleanly to SWE-bench Verified. It also beats Fable 5's 95.0% under the same named evaluation.

But two tempting comparisons from the launch material do not survive a protocol check. Opus 5's 70.6% on OSWorld 2.0 cannot be compared with Fable 5's 85.0% on OSWorld-Verified as though the larger number wins because those are different versions. FrontierBench is an Anthropic-run setup using mini-SWE-agent and a GKE backend, with Opus 4.8 fallback when a safety classifier refused an Opus 5 or Fable 5 trajectory. We keep the 43.3 result as provider-exact evidence, not a benchmark-native ranking row.

Different harnesses can turn a tidy table into a crooked ruler.

Fable 5 keeps small wins at twice the price

The like-for-like coding rows do not show Opus 5 sweeping Fable 5. They show two models separated by tenths of a point on several hard tests, followed by a twofold price gap.

Table 2 Download CSV

Evaluation or price	Opus 5	Fable 5	Edge
SWE-bench Verified	96.0%	95.0%	Opus 5 by 1.0 point
SWE-bench Pro	79.2%	80.0%	Fable 5 by 0.8 points
FrontierCode 1.1 Main	53.4%	53.5%	Fable 5 by 0.1 points
CursorBench 3.2	70.0%	70.5%	Fable 5 by 0.5 points
API input / output	$5 / $25	$10 / $50	Opus 5 costs half

Cursor reports the highest published effort rows on its public evaluation page. Our current data model maps those settings to each canonical model instead of creating a separate row for every effort level.

Calling this a universal Opus victory would be convenient and wrong. Fable 5 holds the small edge on SWE-bench Pro, which may be the better proxy for long, difficult repository work. It also stays fractionally ahead on FrontierCode and CursorBench. Opus 5 wins the buying decision when those differences do not reproduce on the team's own repositories, because paying twice as much for statistical proximity is a poor default.

Reverse the logic too. If Fable 5 clears a production task that Opus 5 repeatedly misses, the benchmark average does not get a vote. Keep the model that passes the task.

The model profile holds more evidence than the rank

As of August 3, the Claude Opus 5 model profile ranks third among 215 public rows at 82.48 and exposes 62 source-displayable benchmark results. That score is a compact index, not a declaration that Opus 5 is the third-best choice for every workload.

Coverage tells the more useful story. Anthropic's launch added strong exact rows for coding, tool use, life sciences, multilingual work, chart interpretation, and mathematical reasoning. It also left important holes. Our July 31 verification audit found six exact overall-ranking slots and 21 missing slots for Opus 5. More published cards do not automatically create more ranking evidence because private evaluations, incompatible versions, and display-only fields do different jobs.

We also kept seven launch claims out of the public benchmark schema. ARC-AGI-1 had no compatible field when the release landed. Automated behavioral-audit and OSS-Fuzz results use safety-specific or count-based scales. Anthropic published relative improvements for organic chemistry, protein-variant prediction, and its life-sciences suite without every absolute value needed for a comparable row.

That refusal matters. Filling empty cells with a predecessor, a nearby chart, or a relative gain would make the profile look complete while making it less true. Missing cells remain missing.

Opus 4.8 users should migrate, while Fable users should test

Table 3 Download CSV

Current position	Recommended move	Where the recommendation loses
Opus 4.8 API workload	Test Opus 5, then migrate if regressions pass	A task-specific behavior or integration constraint breaks under the new model
Fable 5 for general coding	Test Opus 5 against the same repositories	Fable's SWE-bench Pro-style edge reproduces and pays for its 2× rate
Browser or desktop agent	Evaluate Opus 5 on the exact computer-use loop	OSWorld 2.0 does not predict the site's DOM, permissions, or recovery behavior
Latency-sensitive product	Compare Opus 5 Fast with standard Fable 5	Twice-base Fast pricing erases the nominal Opus price advantage
Security research	Keep a model bake-off instead of choosing from this table	The public rows do not establish an exploit-development winner

For Opus 4.8, the burden of proof has flipped. Staying on the old model now requires a measured reason because base pricing is unchanged and the new verified rows are stronger. For Fable 5, migration depends on whether its narrow advantages matter enough to cover the extra invoice.

Use the model comparison tool to assemble the current rows, then run a private evaluation that holds prompts, tools, effort, token budgets, and retry policy constant. Benchmarks can choose the candidates. Your failure cases choose the deployment.

The next independent rows will settle the argument

Opus 5 arrived with enough evidence to replace Opus 4.8 as the default candidate, but not enough like-for-like evidence to erase Fable 5. We are still waiting for a comparable Terminal-Bench 2.1 result. Independent long-horizon agent runs and production-speed measurements matter more than another provider chart.

We will change the recommendation when those rows show a repeatable Fable advantage that pays for its price, or an Opus regression that the launch suite missed. Until then, start the evaluation with Opus 5 and keep Fable 5 as the expensive control.