Claude Opus 5: Anthropic's New Flagship Bets on Efficiency, Not a Price Cut
On 24 July 2026, Anthropic released Claude Opus 5 — barely eight weeks after Opus 4.8 (28 May 2026), and again a release on a tight cadence. Anthropic positions Opus 5 as its new flagship for complex reasoning, coding, and agent tasks, and describes it as coming close to the frontier performance of Claude Fable 5 — at roughly half the cost per task, according to the vendor. The token price itself stays unchanged. For anyone running long agent loops in production, that shift is the actual news.
What Anthropic states about Opus 5
- Released 24 July 2026, available via Claude API, Claude Pro, and Max from day one.
- Token price unchanged: 5 US dollars per million input tokens, 25 US dollars per million output tokens — identical to Opus 4.8.
- Context window: 1 million tokens (default and maximum), up to 128k output tokens, thinking enabled by default.
- Vendor claim: performance close to Claude Fable 5, but per Anthropic at roughly half the cost per task.
- Positioning: new standard model for Max subscribers, strongest selectable option in Pro.
Where things stood
Until 24 July, Claude Opus 4.8 was Anthropic’s flagship. It had launched at the end of May — just 42 days after Opus 4.7 — and locked the token price at 5 US dollars input and 25 US dollars output per million tokens. Its improvements were mainly a cheaper fast mode and alignment near the Mythos class. Anyone who needed the hardest tasks handled reliably reached for Fable 5 — the most expensive and most capable model in the lineup.
The pricing logic followed a clear pattern: more capability cost more tokens or a more expensive model. The relevant comparison was the raw token price per million in and out.
What now applies
1. Same token price, different sales logic. Opus 5 costs exactly the same per token as Opus 4.8 — 5 US dollars input, 25 US dollars output per million tokens. But Anthropic sells the model not on token price, but on cost per completed task: if Opus 5 uses fewer tokens for the same result, the cost per task drops even though the token rate is unchanged. This efficiency figure comes from Anthropic and has not yet been independently verified.
2. Near-Fable performance as a claim. Anthropic describes Opus 5 as close to the frontier intelligence of Claude Fable 5 — at roughly half the cost per task, per the vendor. That is a vendor claim, not a neutral benchmark. If it holds up in practice, Opus 5 would cover a large share of the work that previously required the more expensive Fable 5.
3. 1M token context as default. Opus 5’s one-million-token context window is both default and maximum, output runs up to 128k tokens, and thinking mode is on by default. For long agent loops and large codebases, that is the practically relevant figure.
Reading
The notable thing about this release is not the capability but the metric Anthropic uses to sell it. Until now, token price was the yardstick between models. Opus 5 shifts the focus to cost per task — a figure that depends heavily on how many tokens a model actually needs for a result. That makes sense for agent workloads, where most cost accrues across many steps. But it is also a figure that is harder to verify independently than a raw token rate: token consumption per task varies with prompt, tooling, and task type.
The release cadence is worth noting too. Opus 4.7, 4.8, and now Opus 5 arrived within a few months. For agencies and teams, that means the current flagship is no longer a stable reference point but a moving target. Anyone who hard-wires workflows to a specific model has to re-tune more often than a year ago.
For the actual cost math, that means the interesting question is no longer “what does a million tokens cost?” but “how many tokens does the model need for my typical task?”. That number can only be measured in your own setup — not derived from the price sheet.
What you can do now
If you run Opus 4.8 or Fable 5: measure token consumption per typical task before and after switching to Opus 5, rather than trusting the efficiency figure. Only then do you know whether the “half the cost per task” lands in your setup.
If you needed Fable 5 for heavy tasks: test whether Opus 5 solves the same tasks at comparable quality. If the claim holds, you can move part of the Fable load to Opus 5 at lower cost.
If your model workflows are hard-wired: expect more point releases at short intervals. Keep model selection configurable in one central place, so a switch doesn’t mean rebuilds across the code.