CLAUDE OPUS 5 — CONFIRMED SPECS (LAUNCHED JULY 24, 2026)
● API ID: claude-opus-5 (pinned, no date suffix)
● Standard pricing: $5/M input · $25/M output (same as Opus 4.8)
● Fast mode: $10/M input · $50/M output · ~2.5× faster (API only, research preview)
● Batch API: $2.50/M input · $12.50/M output (50% off standard)
● Cache hits: $0.50/M tokens
● Context window: 1M tokens
● Max output: 128K tokens synchronous
● Knowledge cutoff: May 2026
● Thinking: ON by default, five-level effort toggle (low/medium/high/max)
● Availability: Claude.ai, Claude Max (new default), Claude Pro (top model), Claude Code, Cowork, API, Amazon Bedrock, Google Vertex AI, Microsoft Foundry
● Frontier-Bench v0.1: 43.3% (Opus 4.8: 18.9%) — vendor harness, see caveats below
● ARC-AGI-3: 30.2% — 20× Opus 4.8, 3× next closest model
What Anthropic Actually Shipped
Anthropic launched Claude Opus 5 on July 24, 2026, and it resets what a top-tier model is supposed to cost. It gets close to the frontier intelligence of Claude Fable 5, for about half the price. The new model costs $5 per million input tokens and $25 per million output tokens, matching the price of its predecessor Opus 4.8 and exactly half of Fable 5's $10/$50. Claude Opus 5 is rolling out across all of Anthropic's platforms, where it takes over as the default on Claude Max and claims the top spot among models offered on Claude Pro.
Two new beta features ship alongside it: mid-conversation tool changes that do not invalidate the prompt cache, and automatic fallbacks that route flagged API requests to another model instead of blocking them. The automatic fallback is a direct response to the criticism — including Satya Nadella's July 16 remarks about Fable 5 being "editorially controlled" — that Claude models refuse too many requests. Instead of a hard refusal, flagged requests now route to another model automatically.
The Pricing Story — "Half the Price" Needs Unpacking
The "half the price" framing requires a careful read. The per-token rates are identical to Opus 4.8. The "half the price" line compares Opus 5 to Fable 5 at $10/$50, and Anthropic's per-task cost comparisons come from benchmark runs where Opus 5 needed fewer tokens or turns to finish the same work. On a like-for-like token basis, Opus 5 and Opus 4.8 bill the same.
What this means in practice: if Opus 5 completes the same task as Fable 5 in fewer tokens — which the benchmark data suggests it does on many tasks — then the effective cost per completed task is lower than Fable 5. But if your workload generates similar token counts on Opus 5 as on Opus 4.8, the cost is unchanged. The efficiency gain is real on the tasks where it applies. Test on your own workloads before assuming the cost reduction applies uniformly.
Benchmarks — The Numbers and the Caveats
| Benchmark |
Opus 5 |
Opus 4.8 |
Fable 5 |
Notes |
| Frontier-Bench v0.1 |
43.3% |
18.9% |
Ahead on some tasks |
Vendor harness, Opus 4.8 fallback on safety refusals |
| ARC-AGI-3 |
30.2% |
~1.5% |
N/A |
3× next closest model — significant |
| OSWorld 2.0 |
SOTA |
— |
— |
Computer use / GUI automation benchmark |
| AutomationBench |
SOTA |
— |
— |
Agentic workflow automation |
The benchmark caveat — read before citing these numbers: Most benchmark figures come from Anthropic's own system card or launch harnesses. Several change materially with the effort setting. Some use tools or compaction. A small number include fallback responses from Claude Opus 4.8 after safety refusals. Claude Opus 5 also does not win every comparison — Fable 5 is still ahead on some coding tests, GPT-5.6 Sol is essentially level on web research, and Anthropic reports a higher hallucination rate than Opus 4.8 on one factual benchmark. These are well-disclosed caveats from Anthropic's own documentation — not criticism, just context.
The Effort Toggle — The Most Useful New Feature
It's a per-request setting — low, medium, or high — that controls how much reasoning effort the model spends. Low is faster and cheaper for routine work; high lets the model think longer on hard problems. It's Anthropic's lever for balancing cost against capability. Opus 5 ships with five levels: low, medium, high, and max effort. Thinking is ON by default. This is the feature no previous Anthropic model shipped with — the ability to dial reasoning intensity per request rather than picking a model for the whole task.
In practice: use low effort for classification, routing, and simple generation. Use high or max effort for the hard problems — complex debugging, architectural decisions, multi-step research. The per-request control means you are not paying max-effort prices for every API call, which is what made Fable 5's $10/$50/M pricing so expensive for mixed workloads.
Opus 5 vs Fable 5 — When to Use Each
| Model |
Input /1M |
Output /1M |
SWE-bench Pro |
Effort control |
Best for |
| Claude Opus 5 |
$5 |
$25 |
Not published |
5-level toggle |
Daily use, mixed workloads, agentic tasks, Claude Max default |
| Claude Fable 5 |
$10 |
$50 |
80.4% #1 |
No toggle |
Production coding requiring published 80.4% SWE-bench accuracy |
Opus 5 is built for daily use, not for saving up for hard problems. It runs more efficiently than other models, which is why Anthropic made it the default on Max. The practical guidance: most teams should default to Opus 5 for their Claude workloads. Switch to Fable 5 only for production code changes where the published 80.4% SWE-bench Pro accuracy specifically matters — and where you can justify the 2× cost premium. Fable 5 remains in Claude Max at 50% weekly limits as before.
Fast Mode — What It Is and When to Use It
Fast mode doubles the rate ($10/$50) and runs about 2.5× faster. Fast mode is currently a research preview available on the Claude API only — not on claude.ai or in subscriptions. At $10/$50/M it costs the same as Fable 5 standard mode, but runs 2.5× faster. The use case is latency-sensitive production applications where you need frontier-class responses in real time — not cost-optimised batch processing. For most users, standard Opus 5 at $5/$25/M is the right default. Fast mode is for the specific case where speed matters more than cost.
What Changes on Claude Max and Claude Pro
Claude Max ($100/mo): Opus 5 is now the default model. Fable 5 remains available at 50% weekly limits. If you were on Claude Max before July 24, you are now on Opus 5 without changing any settings. Your Fable 5 credits are unchanged.
Claude Pro ($20/mo): Opus 5 is now the strongest model available to Pro subscribers. This is a significant upgrade — Pro subscribers were previously limited to Sonnet 5 and Haiku as the primary models. Opus 5 on Pro is a meaningful capability increase at the same subscription price.
Migration note for API users: The model ID is claude-opus-5. Do not use claude-opus-4-8 as a drop-in — while pricing is the same, the model behaviour differs significantly (Opus 5 has thinking ON by default, the effort toggle, and the automatic fallback behaviour). Test before migrating production workloads.
The Automatic Fallback — A Direct Response to the Refusal Problem
Automatic fallbacks route flagged API requests to another model instead of blocking them. This is a direct response to the criticism from Satya Nadella and others that Claude models refuse too many requests. Instead of a hard refusal that breaks an application, flagged requests now automatically route to another model — the application gets a response, not an error. The Anthropic system card notes that Opus 4.8 served as the safety fallback in benchmark runs for both Opus 5 and Fable 5, which explains why some benchmark scores include Opus 4.8 responses. In production, the fallback model will be configurable. This is the right architecture for enterprise applications that cannot tolerate refusal-induced failures.
Sources: Anthropic official announcement · Kingy.ai · AIReiter · ExplainX.ai · DigitalApplied · AlphaCorp.ai · ClaudeFast · QZ · WinCentral · Related: Fable 5 vs GPT-5.6 Sol → · Claude Max vs ChatGPT Pro → · Claude plan limits →