Claude Opus 5 is Anthropic's new default model, released on 24 July 2026. It matches or beats Fable 5 on several agentic and reasoning benchmarks, costs half as much on input tokens, is not subject to Anthropic's 30-day data-retention policy, and ships with an effort dial so teams can tune spend against intelligence. See the official announcement for full specs.
TL;DR
- $5/M input, $25/M output — half Fable 5's input cost, same as Opus 4.8.
- Beats Fable 5 on Frontier-Bench v0.1 (43.3% vs 33.7%) and GDPval-AA v2 (Elo 1,861 vs 1,747); trails narrowly on SWE-bench Pro (79.2% vs 80.0%).
- ~85% fewer safety-classifier trips than Fable 5; no 30-day data retention; new Automatic Fallbacks beta.
- New default on Claude Max; strongest on Claude Pro. Reach for Fable 5 only for true long-horizon autonomy.
What is Claude Opus 5 and why does it matter?
Claude Opus 5 is Anthropic's fourth model release in under two months, following Opus 4.8 in May and Mythos 5, Fable 5 and Sonnet 5 in June. The API identifier is claude-opus-5. On Claude Max it is the new default; on Claude Pro it is the strongest available model. It matters because it collapses a familiar trade-off: teams choosing a frontier model used to pay frontier prices for capability they rarely used. Opus 5 sits close enough to Fable 5 on most tasks that the price gap becomes hard to justify.
Independent write-ups make the same call. TechCrunch frames Opus 5 as cheaper and less restrictive than Fable 5, and likely preferable in most use cases. 9to5Google calls it near-Fable 5 performance for half the price. MacRumors' launch report corroborates the pricing and benchmark story.
How does Claude Opus 5 compare to Fable 5 and GPT-5.6 Sol on benchmarks?
The published scores tell a consistent story: Opus 5 wins on agentic reasoning and knowledge work, and loses narrowly on one raw coding benchmark. An independent benchmark verification table reproduces the headline numbers.
| Benchmark | Opus 5 | Fable 5 | GPT-5.6 Sol |
|---|---|---|---|
| Frontier-Bench v0.1 (agentic coding pass rate) | 43.3% | 33.7% | 34.4% |
| ARC-AGI-3 (novel problem solving) | 30.2% | — | 7.8% |
| GDPval-AA v2 (knowledge work, Elo) | 1,861 | 1,747 | 1,736 |
| SWE-bench Pro | 79.2% | 80.0% | — |
Two numbers stand out. On ARC-AGI-3, Opus 5's 30.2% is roughly three times the next-best score — a large gap on a test designed to resist memorisation. On Frontier-Bench, Opus 5 beats both flagship rivals on tasks that need tool use and iteration. Anthropic says Opus 5 was "much stronger at verifying its work and iterating carefully until it succeeds" — the launch post describes it solving a Frontier-Bench task with no drawing view by writing its own computer-vision pipeline to extract geometry from raw pixels and rebuild the part. On real integrations the picture holds: Cursor reports Opus 5 within 0.5% of Fable 5's peak on CursorBench 3.2 at roughly half the cost per task. On OSWorld 2.0 for computer use, Anthropic says Opus 5 beats Fable 5's best result at just over a third of the cost.
How much does Claude Opus 5 cost and how does the effort setting work?
Pricing matches Opus 4.8: $5 per million input tokens and $25 per million output tokens — half of Fable 5's input price. A fast mode is available at ~2.5x default speed for 2x the base price, useful for interactive UI work. The new lever is the effort parameter (low, medium, high, xhigh, max): higher effort produces more internal deliberation for harder problems; lower effort conserves tokens on routine calls. In practice, set a default in your API client and override per call for root-cause analysis or multi-file refactors that demand deeper reasoning.
For orchestration tips that pair well with these settings, see our free Claude Code 2026 setup guide.
What are the safety and data-retention changes?
Two operational changes matter as much as the benchmarks:
- No 30-day retention requirement. Fable 5 and Mythos 5 sit under Anthropic's 30-day data-retention policy; Opus 5 does not. For teams under stricter data-handling contracts, this alone can decide the model choice.
- ~85% fewer safety-classifier interventions. Anthropic expects classifier trips on Opus 5 to fire ~85% less often than on Fable 5. The policy on sensitive categories is unchanged: source-code vulnerability search is allowed; binary-based vulnerability scanning, penetration testing and exploit generation remain blocked. Anthropic also calls Opus 5 its most aligned model to date on automated behavioural audits. Automatic Fallbacks (opt-in beta) complements this: when a classifier trips, the request is routed to a less-powerful model that returns a functional response instead of an error — so one false-positive trip no longer breaks a whole pipeline.
Who is using Claude Opus 5 and what results are they reporting?
Named early-access customers back the "Fable 5 class at half the cost" framing, per quotes in the Anthropic announcement:
- Cognition (Devin). Scott Wu: approaches Fable 5 at half the cost, strong on difficult debugging and root-cause analysis.
- Cursor. Sualeh Asif: near-Fable 5 intelligence at Opus speed and cost, just under Fable 5 on CursorBench.
- Zapier. Wade Foster: topped AutomationBench, ran a full churn-prevention sequence at 100% pass rate.
- Others. AWS Deepak Singh flags it for multi-step agentic coding (now in Kiro); Box +8% overall / +17% due diligence; Lovable +22% on hardest agentic coding.
Anthropic also reports gains over Opus 4.8 across structural biology, organic chemistry and bioinformatics — organic chemistry up 10.2 pp, protein tasks up 7.7 pp.
When should you still use Fable 5?
Fable 5 keeps a narrow edge on SWE-bench Pro (80.0% vs 79.2%) and on long-horizon autonomous agent runs where its raw ceiling matters more than cost. If your workload is a multi-day autonomous agent with irreversible actions, keep Fable 5. For almost everything else — daily coding, retrieval-augmented workflows, structured knowledge tasks, agentic coding, computer use — Opus 5 is the sensible default.
For the wider field, see our Qwen 3.8 vs Fable 5 vs GPT-5.6 Sol comparison, our take on what to use during the Gemini 3.5 Pro delay, the Kimi K3 usage guide, and the Inkling open-weight fine-tune guide.
Limitations
- The 85% reduction in classifier interventions is Anthropic's own expectation, not a measured production figure.
- SWE-bench Pro is one benchmark; pipelines tightly tuned on it may see a small regression on switch.
- Automatic Fallbacks converts a hard error into a quieter, weaker response — log both trips and downgrades.
FAQ
Q: When was Claude Opus 5 released? A: 24 July 2026 — Anthropic's fourth model release in under two months.
Q: How much does Claude Opus 5 cost? A: $5 per million input tokens and $25 per million output tokens, matching Opus 4.8 and half Fable 5's input price. Fast mode runs ~2.5x faster for 2x the base price.
Q: Is Opus 5 better than Fable 5? A: On most published benchmarks yes: Frontier-Bench v0.1 (43.3% vs 33.7%), GDPval-AA v2 (Elo 1,861 vs 1,747) and ARC-AGI-3 (30.2%) all favour Opus 5. Fable 5 keeps a narrow lead on SWE-bench Pro (80.0% vs 79.2%).
Q: What is the effort setting?
A: A parameter accepting low, medium, high, xhigh, max. Higher effort spends more reasoning tokens on hard problems; lower effort conserves cost on routine calls.
Q: Does Opus 5 carry the 30-day data-retention policy? A: No. Fable 5 and Mythos 5 do; Opus 5 does not. Check your specific contract for exceptions.
Q: What are Automatic Fallbacks? A: An opt-in beta that routes a safety-classifier-blocked request to a less-powerful model that returns a functional response instead of an error, so pipelines do not hard-fail on false positives.
Discussion
0 comments