Last updated: July 2026
Key Takeaways
- Claude Opus 5 launched July 24, 2026 at half Fable 5's token rate and is now the default model on Claude Max.
- On cybersecurity and biology, Opus 5 is the less restricted model. Anthropic expects its cyber classifiers to intervene around 85 percent less often than Fable 5's.
- No hosted model can promise it will not hand your request to a different model mid-task, and that includes Kimi K3.
Anthropic released Claude Opus 5 on July 24, 2026. Launch coverage will center on the benchmark table, where Opus 5 takes new state-of-the-art results on Frontier-Bench v0.1 and GDPval-AA while landing within a percentage point of Fable 5 on several coding evaluations. That table answers a question most subscribers do not have.
The practical question is narrower: there is work in front of you right now, so which model do you open, what will each one refuse, and what does the choice cost. The answer moved twice in one week, first on July 20 when Fable 5 access split permanently along plan lines, and again today. Here is the current state, starting with the part that runs backwards from expectation.
The short answer
Use Opus 5 for almost everything. It costs half what Fable 5 costs per token, it is the new default on Claude Max and the strongest model available on Claude Pro, and on Anthropic's own comparison table it leads Fable 5 on agentic terminal coding, knowledge work, novel problem-solving, agentic search, computer use, business workflows, and biology.
Fable 5's remaining leads are narrow, and two of them cannot be bought. Anthropic's table shows Fable 5 ahead by 0.2 points on multidisciplinary reasoning without tools, 0.9 on one agentic coding benchmark, 0.1 on another, and 1.6 on legal agent work. The two larger gaps, health at 66.0 percent and biology at 89.0 percent human-solved, carry a Mythos 5 label rather than a Fable 5 one. Mythos 5 is the unrestricted sibling available only to organizations approved through Anthropic's Project Glasswing, as we covered in our Fable 5 and Mythos 5 specifications breakdown. If you are choosing between two models you can actually subscribe to, those two rows are not on the table.
| What you are doing | Model | Reason |
|---|---|---|
| Drafting, research, general knowledge work | Opus 5 | Leads on GDPval-AA v2 at half the token cost |
| Reviewing source code for vulnerabilities | Opus 5 | Fable 5's classifiers block far more of this work |
| Life sciences research | Opus 5 | Biology requests blocked on Fable 5 now route here |
| Legal agent work, contract redlines | Fable 5 | Only clear purchasable lead: 13.3 against 11.7 |
| One task spanning a million tokens of context | Fable 5 | Carries the 1M window and 128K maximum output |
| High-volume repetitive batch work | Sonnet 5 | Cheapest capable tier during introductory pricing |
| Work that must not be rerouted or logged elsewhere | Open weights, local | No hosted model can make that guarantee |
Benchmark figures are Anthropic's own, published July 24, 2026. Independent replication was not available at publication.
What each model costs as of July 24, 2026
| Model | Per million input | Per million output | Plan position |
|---|---|---|---|
| Opus 5 | 5 US dollars | 25 US dollars | Default on Max, strongest on Pro |
| Fable 5 | 10 US dollars | 50 US dollars | Bundled at 50 percent of limits on Max, metered on Pro |
| Opus 4.8 | 5 US dollars | 25 US dollars | Fallback target for flagged requests |
| Sonnet 5 | 2 US dollars | 10 US dollars | Default on Free and Pro, introductory rate to Aug 31 |
| Kimi K3 | 3 US dollars | 15 US dollars | Separate vendor, hosted only at publication |
Rates are list prices published by each vendor and read on July 24, 2026. Anthropic has changed Fable 5 access terms four times since June 9. Check your usage page before planning around any of this.
The plan structure matters more than the sticker price. Anthropic announced on July 18 that from July 20, Max and Team Premium plans keep Fable 5 bundled at up to 50 percent of weekly usage limits, while Pro and Team Standard moved to metered usage credits at API rates, softened by a one-time 100 dollar credit. Anthropic attributed the change to serving capacity rather than pricing strategy.
The consequence is concrete. On a Pro plan, choosing Fable 5 over Opus 5 now spends money that choosing Opus 5 does not. On a Max plan both are included, but Fable draws against a ceiling that Opus 5 does not have. For a subscriber asking which model to use for a given task, the honest input to that decision is now partly a billing question rather than a capability question.
Does Opus 5 burn tokens as fast as Fable 5?
Per-token price is the wrong unit for this question. Two models can charge the same rate and produce wildly different bills, because they emit different token volumes to finish the same job. The number that matters is cost per completed task.
On that measure Anthropic reports Opus 5 as materially cheaper than Fable 5 for comparable output. On CursorBench 3.2 at maximum effort, Anthropic states Opus 5 finishes within 0.5 percent of Fable 5's peak score at half the cost per task. On OSWorld 2.0, a computer-use benchmark, Anthropic reports Opus 5 surpassing Fable 5's best result at just over a third of the cost. On Zapier's AutomationBench, Anthropic reports Opus 5 passing more tasks at its lowest effort setting than any other model manages at any setting.
That last detail points at the lever readers can actually pull. Opus 5 exposes an effort setting, and Anthropic's own charts show the cost curve bending sharply across it. Several early-access customers reported the same pattern against Opus 4.8 rather than Fable 5: a legal technology firm measured similar quality with 26 percent fewer tokens at maximum reasoning, and a trading firm reported roughly a seventh of the reasoning tokens and under half the latency. If your bills are running hot, the effort setting is the first thing to test, not the model.
What would change this conclusion: every figure above is vendor-reported, measured as cost per task rather than tokens per task, with the effort setting as an obvious confound. An independent evaluator publishing token counts per completed task, at matched effort levels, across both models, would settle it. None had done so at publication.
What Opus 5 will answer that Fable 5 will not
Yes, and this is the part that inverts the usual assumption. The cheaper, lower-tier model is the less restricted one in two documented categories.
Cybersecurity
Anthropic states that Opus 5's cyber classifiers are proportionally less restrictive than Fable 5's, and that it expects them to intervene around 85 percent less often. Opus 5 is permitted to find vulnerabilities in source code. It still blocks binary-based vulnerability scanning, penetration testing, and exploit generation, and it remains substantially behind Mythos 5 at turning a discovered vulnerability into a working exploit. Anthropic says it deliberately avoided training Opus 5 on cyber tasks at all; the improvement came from general capability gains.
For a security-adjacent reader, that is a real change. Source-code review work that Fable 5 would decline is now inside the line on a model that costs half as much.
Biology
Biology requests blocked on Fable 5 now route to Opus 5 rather than Opus 4.8. Anthropic describes Opus 5 as its most capable generally available model for scientific research, while noting it still has meaningful limits on long-running autonomous research and that Mythos 5 remains stronger for that category.
What still gets handed off
Flagged requests on Opus 5 fall back to Opus 4.8 by default in Claude, Claude Code, and Claude Cowork. Automatic fallbacks are now available in beta on the API, letting flagged requests route to another model instead of failing outright. Documentation across surfaces is not fully consistent about which Opus version receives the handoff, so read the model label on the response rather than assuming.
This is where the structural point sits. Anthropic's own footnote on the Frontier-Bench results states that Opus 4.8 served as the fallback on safety-classifier refusals for both Opus 5 and Fable 5 during scoring. The benchmark numbers already contain the classifier. That is the same measurement problem we worked through when independent testers claimed the returning Fable 5 had been degraded, covered in our analysis of whether Claude Fable 5 was nerfed and in our account of the silent safeguards reversal that established visible fallback in the first place. A capability you cannot isolate from a classifier is a capability you cannot audit from outside.
Should you switch to Kimi K3 instead?
Open weights is not the same as ungated
Kimi K3 is frequently suggested as the less-restricted alternative, and the reasoning is usually that it is an open model. That reasoning does not hold while you are calling someone else's endpoint. A hosted API sits behind whatever policy stack its operator runs, and Moonshot's is neither published in the detail Anthropic publishes nor independently audited. The license property of an open-weight model becomes yours at the moment the weights do, and not before.
The weights are not out yet
Moonshot launched K3 on July 16 and 17, 2026 as a hosted-only release through Kimi.com, Kimi Code, and the API, and committed in its own launch post to publishing full model weights by July 27, 2026, alongside a technical report. As of publication that is a promise, not a shipped artifact. No repository, no checksums, and no license file exist. Moonshot's recent releases used a modified MIT license carrying commercial-use conditions, which is a reasonable prior and not a guarantee; read the license when it lands rather than assuming it.
Even after the drop, K3 is not a desktop proposition. It is a 2.8 trillion parameter sparse mixture-of-experts model, and Moonshot's own serving guidance points at distributed infrastructure. We walked through the hardware arithmetic in our Kimi K3 local hardware reality check, and the short version has not changed: new architectures take time to land in llama.cpp, Ollama, and LM Studio, and until they do, downloading the weights buys you a very large folder.
The one thing no hosted model offers
If the binding requirement is that nothing sits between you and the model, no subscription answers it. Not Opus 5, not Fable 5, not K3 on Moonshot's API. Classifiers reroute, prices change, plan terms shift four times in six weeks, and in June a single export-control directive took Fable 5 offline worldwide for eighteen days. None of that is hypothetical risk. It is the last two months.
The honest counterweight is that nothing you run at home matches Opus 5 on the benchmarks above, and pretending otherwise helps nobody. The trade is capability for permanence. If permanence is what you need, our guide to the best local AI models by VRAM tier maps what actually runs at every memory level, and our local AI hardware guide covers what each budget buys. For most people the correct answer is both: a hosted frontier model for hard work, and a local model for anything that must keep running when someone else's policy changes.
Frequently Asked Questions
Is Claude Opus 5 better than Fable 5?
On most public benchmarks Anthropic published at launch, yes, and at half the token price. Fable 5 retains narrow leads on legal agent work, multidisciplinary reasoning without tools, and two agentic coding benchmarks by margins under one point, plus a one-million-token context window Opus 5 does not match.
Do I still need Fable 5 on a Pro plan?
For most work, no. Since July 20, Fable 5 on Pro and Team Standard bills through metered usage credits at 10 dollars per million input tokens and 50 per million output, while Opus 5 is included as the strongest model on the plan. Fable 5 earns its cost on Pro mainly for long-context work and legal tasks.
Can Opus 5 answer cybersecurity questions Fable 5 blocks?
Yes. Anthropic expects Opus 5's cyber classifiers to intervene around 85 percent less often than Fable 5's, and Opus 5 is permitted to find vulnerabilities in source code. Binary-based vulnerability scanning, penetration testing, and exploit generation remain blocked.
Does Opus 5 use fewer tokens than Fable 5?
Anthropic reports lower cost per completed task rather than publishing token counts directly, citing half the per-task cost of Fable 5 on CursorBench 3.2 and just over a third on OSWorld 2.0. Those are vendor figures. The effort setting is the variable that most affects your actual token spend.
Can I run Claude Opus 5 locally?
No. The weights are closed and inference runs only on Anthropic's infrastructure. The nearest equivalent path is running open-weight models through a tool like Ollama on hardware you own, at lower capability but with no external dependency.
Is Kimi K3 a drop-in replacement for Fable 5?
Not today. K3 launched hosted-only on July 16, 2026, with full weights promised by July 27. Until the repository and license exist, K3 is another vendor's API with another vendor's policies, and at 2.8 trillion parameters it will need distributed infrastructure rather than a desktop even after release.
What happens when Opus 5 refuses a request?
In Claude, Claude Code, and Claude Cowork, flagged requests fall back to Opus 4.8 by default rather than failing. On the API, automatic fallbacks are available in beta. Check the model label on the response, since documentation across surfaces is not consistent about the exact fallback version.
