What happened
Anthropic just released Claude Opus 5, a flagship AI model that Anthropic says delivers near-Fable 5 performance at half the token price. The company is positioning Opus 5 as the smarter, cheaper option for coding and knowledge work, and it's now the default model on Claude Max and the most capable model available on Claude Pro.
The headline numbers are hard to ignore. On Frontier-Bench v0.1, a test of agentic terminal coding, Opus 5 scores 43.3 percent — beating Claude Fable 5 (33.7 percent), GPT-5.6 Sol (34.4 percent), and its own predecessor Opus 4.8 (21.1 percent) by wide margins. On ARC-AGI-3, a benchmark for novel problem-solving, Opus 5 hits 30.2 percent, nearly four times higher than GPT-5.6 Sol. On the knowledge-work benchmark GDPval-AA v2, Opus 5 posts an Elo score of 1,861, ahead of Fable 5 (1,747) and GPT-5.6 Sol (1,736).
### Pricing stays put
Despite the performance jump, Anthropic didn't touch the price. Opus 5 still costs $5 per million input tokens and $25 per million output tokens — identical to Opus 4.8 — and it keeps the 1 million-token context window. A new Fast Mode option runs 2.5x faster but doubles the price, useful when latency matters more than cost.
Why it matters
The timing isn't a coincidence. Anthropic is under real pricing pressure from GPT-5.6 Sol and a wave of low-cost Chinese models, and Opus 5 is a direct answer: match or beat the competition on benchmarks while staying meaningfully cheaper than Anthropic's own top-tier Fable 5. For a market that increasingly treats AI coding assistants as a line item, that price-performance gap is the story.
### The token-efficiency catch
Raw per-token pricing doesn't tell the whole story. Anthropic's own history shows why: Opus 4.7 ended up costing 30 to 40 percent more per task than Opus 4.6, even though both had identical base rates, because the newer model used more tokens to reach an answer. A similar pattern has shown up with Claude Sonnet 5. In other words, a lower sticker price doesn't guarantee a lower bill — it depends on how efficiently the model reasons through a task, not just what it charges per million tokens.
Opus 5 also isn't uniformly ahead. On the DeepSWE v1.1 agentic coding benchmark, GPT-5.6 Sol leads with 72.7 percent, followed by Fable 5 (69.7 percent) and Opus 5 (68.8 percent). Fable 5 and the limited-availability Mythos 5 model still outperform Opus 5 on health and legal-domain tasks. So Opus 5's strength is concentrated in coding and general knowledge work, not every category.
How to use it today
Opus 5 ships with five effort settings — low, medium, high, xhigh, and max — letting users trade accuracy for speed and token spend. Anthropic recommends leaning on "low" and "medium" for everyday tasks, saying they now deliver strong results at a fraction of the cost and latency of earlier Opus models at the same settings. For coding and agentic workflows specifically, Anthropic still recommends starting at "xhigh."
### When more effort backfires
Here's a genuinely useful detail for anyone tuning API costs: on two benchmarks — Frontier-Bench v0.1 and the Artificial Analysis Coding Agent Index — Opus 5 actually scores slightly worse at the "max" effort setting than at the second-highest tier, despite max mode costing more. That means blindly cranking effort to maximum can waste money without improving output. It's worth testing your specific workload at "high" or "xhigh" before assuming "max" is better.
For smaller teams and solo builders who want to experiment with AI models before committing budget to a paid API plan, it's worth testing prompts and workflows on free tools first. Sites like [mykreatool.com](https://mykreatool.com) offer free AI utilities that let you prototype ideas — writing prompts, generating content, or testing automation logic — before scaling up to a metered model like Opus 5.
Who benefits
Developers and engineering teams running agentic coding workflows are the clearest winners. Opus 5's lead on Frontier-Bench v0.1 and its top score on the Artificial Analysis Coding Agent Index (narrowly ahead of GPT-5.6 Sol) suggest it's currently one of the strongest models for autonomous coding tasks — writing, testing, and iterating on code with minimal human intervention. Anthropic says Opus 5 can check and improve its own work through iteration and build its own tools through code when needed, which matters for teams running long, unsupervised coding agents.
Knowledge workers and consultants also stand to gain. The GDPval-AA v2 results suggest Opus 5 handles general professional tasks — research, analysis, drafting — better than Fable 5 or GPT-5.6 Sol, at less than half Fable 5's token cost. Startups and agencies that were priced out of Fable 5 for high-volume use cases now have a cheaper model that closes most of the performance gap.
Risks
The biggest risk is cost unpredictability. Because token efficiency varies by task and effort setting, teams that migrate to Opus 5 without benchmarking their own workloads could end up paying more than expected — the same trap that hit Opus 4.7 and Sonnet 5 users. Anyone switching models should run a side-by-side cost comparison on real production prompts, not just trust the headline per-token rate.
Domain mismatch is the second risk. Opus 5 is not the strongest choice for health or legal work, where Fable 5 and Mythos 5 still lead. Teams in regulated or specialized industries should validate accuracy on their specific domain before switching models purely for cost savings. And the "max" effort setting's inconsistent payoff is a reminder that more expensive doesn't always mean more accurate — always test before defaulting to the priciest option.
Conclusion
Claude Opus 5 gives Anthropic a genuine answer to GPT-5.6 Sol and cheaper Chinese competitors: benchmark-leading coding and knowledge-work scores at the same $5/$25 per-million-token pricing as its predecessor, and roughly half of Fable 5's rate. The catch is that per-token pricing isn't the whole cost picture — token efficiency, effort settings, and task domain all swing the real bill. For coding-heavy teams and knowledge workers, Opus 5 looks like the best value in Anthropic's lineup right now; for health, legal, or highly specialized work, Fable 5 still holds the edge.
Comments 0