Opus 5 Is Half the Price of Fable 5. Turn On Fast Mode and It Isn't.

Anthropic shipped Claude Opus 5 yesterday, and the pitch is clean: near-Fable-5 intelligence at Opus 4.8 prices, which is to say half of what Fable 5 costs. That's a genuine number, not a rounding trick, and it's worth sitting with for a second before the caveat, because the caveat is the part nobody is putting in the headline.

The number that's actually true

Opus 5 launches at $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8. Fable 5, which Anthropic shipped in June, charges $10 and $50. So the base-rate math holds: on Anthropic's own CursorBench 3.2, run at maximum effort, Opus 5 lands within 0.5% of Fable 5's peak score while costing half as much per task. On OSWorld 2.0 it clears Fable 5's best result at just over a third of the cost. Those are vendor numbers, and we'll say more about that in a minute, but the pricing side of the claim is contractual, not marketing copy: it's what the API bills you.

The efficiency claims layered on top are the more interesting part. Anthropic says Opus 5 matches Opus 4.8 on a trading benchmark using roughly a seventh of the reasoning tokens and under half the latency, and that on a legal-document task it holds performance while generating 26% fewer tokens on average. If even half of that survives contact with real workloads, it's a bigger deal than the sticker price, because token count is what actually shows up on your invoice, not the listed per-million rate.

The tier the press release doesn't lead with

Opus 5 also ships with a Fast mode: about 2.5x the default speed, at twice Opus 5's base price. Do the arithmetic and Fast mode lands at $10 input / $50 output per million tokens. Those are not similar numbers to Fable 5's rate. They are the same numbers, to the dollar.

That's not a scandal. Charging more for lower latency is a completely defensible product decision, and plenty of workloads (agents waiting on a tool call, anything user-facing) will happily pay it. But it does mean the "half the price" framing is a claim about the default tier, specifically, not about Opus 5 as a product. The moment a team needs Fable-5-class responsiveness rather than Fable-5-class accuracy, the discount Anthropic is advertising evaporates completely. You end up paying Fable 5 money for an Opus 5 label.

Anthropic is responding to pricing pressure from GPT-5.6 Sol and Chinese competitors. Its new Opus 5 model is designed to close the price-performance gap with the much pricier Fable 5. - the-decoder.com

Why the timing isn't incidental

Anthropic is reportedly preparing an IPO later this year, and a launch that reads as "we cut our flagship's effective price in half" is a good headline to be carrying into a roadshow. It signals aggressive competitive positioning against GPT-5.6 and the wave of Chinese open-weight releases (Kimi K3, DeepSeek V4, GLM-5.2) that have been undercutting everyone on cost this month. None of that makes the release cynical. It does mean the pricing structure deserves the same scrutiny we'd give any other pre-IPO metric: what's the headline number measuring, and what's the number that applies once you actually use the product the way most paying customers will.

This is also the fourth or fifth time this year we've seen an "efficiency" claim (GPT-5.6, Grok 4.5, and now Opus 5) that reduces to a private benchmark, a specific effort setting, and no shared yardstick against the others. That pattern is exactly why we don't run Anthropic's numbers as our headline here. We queue the model into the arena, run it blind against Fable 5, Sonnet, and whatever GPT-5.6 tier is live that week, and let votes accumulate on the same one-shot prompts everyone else answers.

• Opus 5 base: $5 input / $25 output per million tokens, same as Opus 4.8
• Opus 5 Fast: $10 input / $50 output, double the base rate for roughly 2.5x speed
• Fable 5 default: $10 input / $50 output, identical to Opus 5 Fast
• Anthropic's own benchmarks (CursorBench 3.2, Frontier-Bench v0.1, OSWorld 2.0) are self-reported, at stated effort levels, and not directly comparable to what we log

If you want to see how Opus 5 actually performs against Fable 5 rather than take Anthropic's word for the gap, open the coding arena once both are live there and read the votes yourself, not a slide from a launch deck. We covered the Opus/Fable relationship before Opus 5 existed in our look at Claude Mythos vs Claude Fable 5, and the same caution applies here: a model family sharing a lineage doesn't mean shared pricing logic, and Anthropic just proved that by pricing two tiers of the same model $5 apart on input and $25 apart on output.

What to actually do with this

If your workload tolerates default latency, Opus 5 at $5/$25 is a straightforward win over Fable 5 on cost, assuming the quality gap really is as small as CursorBench suggests. If your workload needs Fast mode, run the comparison again before switching, because you are no longer choosing a cheaper model, you're choosing a differently priced one. The mistake to avoid is reading one press cycle's "half the price" framing and assuming it survives whichever tier your actual product runs on. It doesn't, not automatically, and the fine print is one multiplication away from the headline.

For readers following this in Polish, nowosci.ai has been tracking the same release cycle and is worth a look for the regional angle on how European enterprise buyers are reacting to the Opus 5 pricing move.