Anthropic released Opus 5 today at $5 per million input tokens and $25 per million output tokens, holding pricing flat versus its predecessor while delivering performance at or slightly ahead of the company's more expensive Fable model. On Anthropic's own charts, Opus 5 edges past Opus 4.8 and OpenAI's GPT-5.6-Sol across Frontier-Bench, DeepSWE, and adjacent coding evaluations. The pitch is not a capability breakthrough. It is roughly Fable-class output at approximately half the cost of Fable.
That framing matters because the last Opus update — 4.5 — was genuinely a step change in agentic coding. Opus 5 is not that. Anthropic's benchmark deltas over Opus 4.8 are iterative, not radical, and the company is explicit that Mythos remains the flagship for the hardest tasks. Opus 5 is positioned one rung down: not the frontier, but close enough for most software work at materially lower spend.
The cybersecurity gap is deliberate. Anthropic did not give Opus 5 cutting-edge training on offensive security tasks, and the model lags well behind both Fable and Mythos on vulnerability exploitation. In Anthropic's words, Opus 5 remains "relatively good at finding cybersecurity vulnerabilities" but is "substantially behind Mythos 5 on the exploitation of those vulnerabilities." As a consequence, Opus 5 also sheds some of Fable's more contested guardrails, including the 30-day data review retention policy that drew scrutiny at Fable's launch.
“Opus 5 isn't as remarkable as Mythos, but it's more widely available and much more affordable.”— Samuel Axon, Senior Editor, Ars Technica
Key facts
- 01Opus 5 is priced at $5 per million input tokens and $25 per million output tokens, roughly half the cost of Anthropic's Fable model.
- 02On Frontier-Bench and DeepSWE, Opus 5 lands at or slightly ahead of Fable and beats Opus 4.8 and OpenAI's GPT-5.6-Sol on most tasks.
- 03Anthropic deliberately withheld cutting-edge cybersecurity training, leaving Opus 5 substantially behind Mythos 5 on vulnerability exploitation.
- 04Kimi K3, a Chinese open-weight model, offers similar performance at $15 per million output tokens — undercutting Opus 5.
- 05Opus 5 drops Fable's 30-day data review retention policy, reflecting a lighter safety posture on the cheaper tier.
The competitive backdrop is what makes the pricing decision interesting. Kimi K3, the recently released Chinese open-weight model, delivers comparable performance at $15 per million output tokens — 40% cheaper than Opus 5 on the output side, and running on infrastructure customers can host themselves. Open-weight and locally hosted models have moved from novelty to genuine procurement option for engineering leaders trying to bring inference costs under control.
That pressure is showing up in tooling. Cursor and Meta have both built model routers that automatically dispatch prompts to different models based on task complexity, sending cheap requests to smaller or open models and reserving frontier models for hard problems. The economic logic is straightforward: if 70% of coding requests can be handled by a mid-tier or open-weight model, paying frontier prices for every token is waste. Routers turn that logic into infrastructure.
Anthropic's response is to make the mid-tier itself more attractive. Opus 5 is priced to be the model a router picks by default — cheaper than Fable, more capable than most open alternatives, and consistent with Anthropic's safety posture. Holding the price flat while shipping better benchmarks is functionally a discount, and it targets exactly the segment where routers are pulling volume away from frontier tiers.
The trade-off is real. Anthropic has recently become a popular pick for coding and software development work, and Opus is now the model line carrying most of that demand. Concentrating on token efficiency rather than capability means Anthropic is competing on the same axis as the open-weight ecosystem, where price floors keep falling. Kimi K3 already undercuts Opus 5 on output pricing. If open-weight quality continues to close the gap on coding benchmarks, the mid-tier squeeze gets tighter, not looser.
There is also a benchmark-versus-reality question. Anthropic's charts are Anthropic's charts. Independent evaluations on Frontier-Bench and DeepSWE will take weeks to settle, and Opus 5's real edge against Fable at half the cost has to survive third-party testing before enterprise buyers rearrange their routing rules. The security-training gap versus Mythos is also a genuine ceiling for any team using Opus in offensive security or advanced red-team workflows — those users stay on Mythos.
The Opus 5 launch is a signal about the shape of the frontier model business in 2026. Capability leaps between adjacent versions are getting smaller, cost is getting louder, and routers mean model providers no longer get to charge frontier prices for tasks the frontier isn't needed for. Anthropic is choosing to compete on efficiency at the mid-tier while keeping Mythos as the capability halo. That is a defensible posture, but it puts Opus in direct combat with open-weight models on price — a fight where the trajectory has been running one direction for a year, and it isn't Anthropic's.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
