Anthropic's new Opus comes within striking distance of Fable 5, the tier above it, while charging the same $5 / $25 as Opus 4.8. It leads coding, agentic automation, computer use, and abstract reasoning, and Anthropic calls it their most aligned model to date. For most of the hard work we build, this is the new default.
Standard pricing is $5 input and $25 output per million tokens, identical to Opus 4.8. But Opus 5 now competes with Fable 5, which costs twice as much. The price did not move; the ceiling did.
On CursorBench 3.2 at max effort, Opus 5 lands within half a percent of Fable 5's peak while costing half as much per token. On OSWorld 2.0 computer use it surpasses Fable 5's best at roughly a third of the cost.
Anthropic reports Opus 5 more than doubles Opus 4.8's score on Frontier-Bench v0.1 and sets a new state of the art. On ARC-AGI 3 abstract reasoning it is about three times the next-best model.
The behavioural audit scores misaligned behaviour at 2.3, the lowest they have measured, with the lowest deceptive-behaviour rates yet. Opus 5 is also much stronger at verifying its own work and iterating before it reports back.
You can add or swap tools while a conversation is running (beta), and the API can automatically route flagged calls to a fallback model (beta). Both remove glue code from long agent runs.
From Anthropic's launch numbers, plus the ratios they published where they did not give an absolute score. The pattern is consistent: Opus 5 clears Opus 4.8 by a wide margin, runs level with Fable 5 (the tier above), and does it at half Fable 5's price.
| Capability | Opus 5 | Opus 4.8 | Fable 5 | GPT-5.6 Sol |
|---|---|---|---|---|
Agentic coding SWE-Bench Pro | 79.6% +10.4 pts vs 4.8 | 69.2% | 80.3% | 74.1% |
Frontier coding Frontier-Bench v0.1 more than 2x Opus 4.8; leads all | 71.8% +38.3 pts vs 4.8 | 33.5% | 68.4% | 65.9% |
Coding, max effort CursorBench 3.2 within 0.5% of Fable 5, half the price | 79.9% +8.7 pts vs 4.8 | 71.2% | 80.3% | 77.5% |
Abstract reasoning ARC-AGI 3 about 3x the next-best model | 45.6% +32.8 pts vs 4.8 | 12.8% | 15.1% | 14.9% |
Agentic automation Zapier AutomationBench about 1.5x the next-best pass rate | 62.4% +22.9 pts vs 4.8 | 39.5% | 41.8% | 42.1% |
Agentic computer use OSWorld 2.0 beats Fable 5 at about a third of the cost | 61.2% +8.8 pts vs 4.8 | 52.4% | 58.7% | 55.9% |
Knowledge work GDPval-AA | 2040 +150 vs 4.8 | 1890 | 1985 | 1802 |
Science Organic chemistry +10.2 pts vs 4.8, as Anthropic reports | 74.8% +10.2 pts vs 4.8 | 64.6% | 72.0% | 69.1% |
Anthropic published several of these as relative claims rather than raw scores: more than 2x Opus 4.8 on Frontier-Bench, within 0.5% of Fable 5 on CursorBench at half the cost, about 3x the next-best on ARC-AGI 3, and a +10.2 point organic-chemistry gain over 4.8. We render them directionally and re-run our own evals on customer tasks before recommending a switch.
What changes for the engineering team. Two comparisons that matter: Opus 5 against the Opus it replaces (4.8), and against Sonnet 5, where the cheaper, faster tier still clears routine work. The headline shift is above, not below: Opus 5 removes many of the reasons to pay Fable 5 rates.
| Dimension | vs Opus 4.8 | vs latest Sonnet |
|---|---|---|
Frontier coding | More than doubles Opus 4.8 on Frontier-Bench v0.1 and adds about 10 points on SWE-Bench Pro. It now rivals Fable 5, so a lot of work that needed the premium tier no longer does. | Sonnet 5 still handles most local edits and review cheaply. Reach for Opus 5 on architecture, migrations, and multi-file refactors. |
Cost against the frontier | Identical standard pricing to Opus 4.8 ($5 / $25). The capability went up, the sticker did not. | Opus 5 costs more per token than Sonnet 5, and it earns it when one Opus call replaces several Sonnet retries. The new lever is above it: Opus 5 matches or beats Fable 5 on several tasks at half or a third of the cost. |
Agent reliability | Stronger at verifying its own work and iterating before reporting back, and Anthropic's most aligned model yet. Those are exactly the failure modes that bite on long agent runs. | Route the easy, high-volume steps to Sonnet 5 and the judgement-heavy ones to Opus 5, behind one abstraction and one eval suite. |
Migration risk | Behind a vendor-neutral abstraction the switch is a config change plus an eval pass. Same price, so there is no budget reason to wait. Most prompts work identically. | Different positioning, not a replacement. We run both and route by task difficulty at runtime. |
Inside a Kensink build, model selection is a routing decision the agent makes at runtime, not a vendor commitment frozen at design time. Opus 5 mostly changes which calls still justify Fable 5.
The reframing that matters. Opus 5 lands within 0.5% of Fable 5 on coding and beats it on some computer-use tasks, at half or a third of the cost, while holding the $5 / $25 price of Opus 4.8. It is now the default tier for hard work, with Fable 5 reserved for the residual hardest cases.
Add, swap, or remove tools while a conversation is running (beta on the Claude Platform). Permissions and available actions can change as an agent works, without restarting the conversation or rebuilding the prompt.
The API can automatically route flagged calls to an alternative model (beta). Useful as a safety net inside a long agent run, though we still make the primary routing decision explicitly behind our own abstraction.
Anthropic highlights best-in-class animations, games, and 3D work, alongside gains in financial modelling, legal document review and redlining, and code review and debugging. Opus 5 also builds and runs its own verification harnesses without external prompting.
Opus 5 stays behind Mythos 5 on cybersecurity and OSS-Fuzz exploit development, and Anthropic says it does not advance the dual-use frontier. Enterprises with sensitive cyber needs get a path through the Cyber Verification Program.
The behavioural audit scores misaligned behaviour at 2.3, the lowest Anthropic has recorded, with the lowest deceptive-behaviour rates they have measured. On long autonomous runs, deception and quiet cooperation with misuse are the failure modes that cost the most, so this is the improvement that matters.
Anthropic reports Opus 5 is much stronger at verifying outputs and iterating carefully, and demonstrates it building custom test harnesses without external validation. That is the difference between an agent that is confidently wrong and one that catches its own mistake before it ships.
Opus 5 remains behind Mythos 5 on cybersecurity tasks and on OSS-Fuzz exploit development. Anthropic gates the higher-risk class separately and offers the Cyber Verification Program for enterprises that need sanctioned access for defensive work.
Same price as 4.8, materially stronger, and it closes most of the gap to Fable 5. Behind our vendor-neutral abstraction the switch is a config change plus an eval pass. There is no commercial reason to wait.
The reasons to reach for the premium tier were frontier coding and computer use. Opus 5 now matches or beats Fable 5 on several of those at half or a third of the cost. We route to Fable 5 only for the residual hardest cases, and re-check that boundary against our own evals.
Self-verification and lower deception are exactly what fails silently over hours of autonomous work: translation runs, deep research, migrations, slide-building. This is where Opus 5 earns its keep, not on single-shot prompts.
Mid-conversation tool changes and automatic fallbacks are genuinely useful for long-running agents. They do not replace task-specific evals on your data. We run the customer eval suite on every model change, every release.
Every model we integrate runs through the same operating system. Three pillars, sixteen layers, one Compound Growth Loop. The methodology that keeps AI work from rotting after the first ship.
Read the K-FrameworkDirect API integration with the model. No LangChain, no orchestration vendor, no agent framework built on quicksand. Typed contracts, the same way we wire up Postgres.
An eval suite built from your real tasks gates every prompt and model change. Quality is measured before it ships, not vibed in a demo.
Governance, audit, and oversight wired in from day one. Who called what, with which prompt version, at what cost. Your auditors get answers, not screenshots.
A model in production without observability is roulette. We instrument every integration so engineering and finance can see the same numbers, and so a regression at 3am surfaces before a customer opens a ticket.
Tokens in, tokens out, dollars spent. Sliced by feature, tenant, and route. Budgets enforced where it matters.
Real distributions, not averages. We know which routes are slow, and why.
The same eval suite that gates a release runs continuously in production. A regression on real traffic surfaces fast.
PII scrubbed at the proxy, shipped to your SIEM. Retention controls match your compliance window.
Dashboards your team owns, not ours. At handoff you get the queries, the alerts, and the runbook. We are not in the path to read your metrics.
