Anthropic shipped Claude Opus 5 on July 24. Four days later I'm still sitting with the benchmark numbers, because they tell a story I didn't expect: on the metrics that actually matter for builder work, Opus 5 outperforms Fable 5 at exactly half the token price.
That's not "close enough." That's a genuine inversion.
What the Numbers Actually Say
Fable 5 has been Anthropic's flagship since June. It costs $10 per million input tokens, $50 per million output tokens. Opus 5 costs $5/$25 per million tokens, which is identical to Opus 4.8. Same price point, better performance on the key evals.
On Frontier-Bench v0.1, the agentic coding benchmark I weight most heavily right now, Opus 5 scores 43.3%. Fable 5 scores 33.7%. On SWE-bench Pro it posts 79.2%. On ARC-AGI-3, which tests novel problem-solving rather than memorized patterns, Opus 5 scores 30.2%. The next-best model scored 7.8%. Fable 5 wasn't even tested on it.
Artificial Analysis pegged the weighted average cost-per-task at $2.03 for Opus 5 versus $2.75 for Fable 5. Cheaper per token AND cheaper per task AND better on the hardest agentic and reasoning evals. That's a genuinely strange situation.
The Effort Mode Change You Need to Know
In the 4.x era, Anthropic told you to start at xhigh effort for agentic work and dial down if cost was a concern. Opus 5 flips that guidance. You start at high and only reach up if you need it.
This matters in practice. At low effort, Opus 5 delivers well over half its peak score using roughly a third of the output tokens. If you're running a high-volume agentic loop where most steps are fairly routine, low or medium effort will be significantly cheaper and still produce useful work. There's also a fast mode that runs about 2.5x faster at $10/$50 per million tokens, which makes sense when latency is the constraint rather than cost.
The practical implication: for the "thinking budget" pattern I've been using in my agents (where each tool call gets a budget of thinking tokens scaled to how hard the step is), Opus 5 maps cleanly to low/medium/high based on step complexity. You're not manually tuning thinking_budget integers anymore. You pick a tier.
When Would You Still Use Fable 5?
Honestly, the case is thinner than it was three weeks ago. The main argument I can construct is that your team has specific prompts or workflows already tuned to Fable 5's output style, and the cost of re-testing is real. Anthropic isn't retiring Fable 5, so nothing breaks.
There's also a practical privacy difference: Fable 5 requires a 30-day prompt retention agreement for general API access. Opus 5 has no data retention requirement. For any enterprise use case with PII in the prompt, that's a cleaner posture by default.
If Fable 5 still wins on a specific narrow benchmark that's critical to your use case, that's a valid call. But "we use Fable because it's the best" is harder to defend now that the benchmarks show otherwise on the tasks most builders actually run.
How I'm Rethinking My Routing
I currently route between Sonnet 5 for fast classification and extraction tasks, Fable 5 for deep reasoning and multi-step planning, and Sonnet 5 with extended thinking as a middle tier. I'm collapsing that to Sonnet 5 and Opus 5, with effort level doing the work that model selection used to do.
Sonnet 5 for tasks that don't need reasoning across a large context or complex planning. Opus 5 at low or medium effort for high-volume tasks that benefit from some thinking. Opus 5 at high effort for anything where quality matters more than throughput, including the orchestration and planning steps in multi-agent workflows.
I haven't run this in production yet. I've done enough spot checks to feel confident in the direction. The real calibration will come from watching actual task success rates, not benchmark extrapolation. One thing I'm watching closely: whether effort level interacts cleanly with the types of tool calls my agents make. Early impression is that medium handles most real agentic steps better than the benchmark numbers alone would suggest.
The Inference Economy Is Moving Faster Than Most Teams Realize
Anthropic's fourth Claude 5 release in under two months has pushed the frontier cheaper in a way that compounds. Sonnet 5 was already near-Fable-4.8 quality at a fraction of the price. Opus 5 now sits within a single point of Fable 5 on Artificial Analysis's intelligence index, at half the token cost, and ahead on the hardest evals.
If you're defaulting to the most expensive model in your stack out of habit rather than demonstrated need, now is a good time to run the audit. The math has shifted significantly, and it'll shift again. Build your routing layer to be model-agnostic and parameterized rather than hardcoded to a specific model name. That's the only real hedge in a landscape that's moving this fast.
Comments
Post a Comment