← The Polylog AI Intelligence Brief
Morning Edition · Saturday, October 17, 2026Published at 2:12 AM EDT · New York
Anthropic prices Claude Haiku 5.5 at one-tenth of Haiku 4.5 rates for prompts under 100,000 tokens
Prompts above that length cost five times more and a newer tokenizer adds about 30% more tokens, so Anthropic puts the average saving near 75%, and it reports a computer-use score of 72.4% on a subset of OSWorld 2.1.
Anthropic launched Claude Haiku 5.5 on October 7 and calls it the fastest and cheapest small model it has released. For prompts up to 100,000 tokens, it charges $0.10 per million input tokens and $0.50 per million output tokens, thinking tokens included, according to eesel AI's breakdown. Cache reads cost $0.01, and batch rates are half the standard rates. Haiku 4.5 cost $1 for input and $5 for output, so the short-prompt tier is one-tenth of the old price, PPC Land reported.
Two details reduce the size of the headline cut. Prompts above 100,000 tokens cost $0.50 for input and $2.50 for output. Haiku 5.5 also uses the tokenizer introduced with Claude 4.7, which produces about 30% more tokens for the same text. Anthropic says about 90% of Haiku 4.5 requests fell below the threshold, and it estimates the average saving at around 75% once the request mix and tokenizer are factored in, AI Weekly reported. Engineers running retrieval-heavy or long-document pipelines need to measure their own prompt-length distribution, because jobs that cross the threshold pay the higher tier.
Anthropic's reported capability gain is large. On the offline subset of the OSWorld 2.1 computer-use benchmark, Haiku 5.5 scores 72.4% against 15.7% for Haiku 4.5. It scores 64.8% on SWE-Bench Pro. These are Anthropic's numbers on Anthropic's chosen subsets, and outside reproduction has not appeared. The release follows Claude Sonnet 5.5 on September 28, which Anthropic says runs 30% faster and costs up to 30% less than Sonnet 5 for most work.
Other vendors have already moved on price. Meta opened its Muse Spark 1.1 API at $1.25 per million input tokens and $4.25 per million output tokens. Mark Zuckerberg, Meta's chief executive, described that as about a quarter of what Anthropic and OpenAI charge for comparable models, Implicator reported. Anthropic is now cutting prices at the bottom of its lineup, where open-weight models and hyperscaler-hosted alternatives compete most directly.
What this means
Anthropic is giving up margin on high-volume small-model traffic to keep agent workloads from moving to open-weight models served on commodity GPUs. It chose a pricing structure that protects revenue on long-context calls, which cost the most to serve. Inference resellers and small-model API vendors that priced against Haiku 4.5 are exposed, because a tenfold cut in list price resets the benchmark buyers compare against. Investors who value labs on API revenue growth should expect volume to rise faster than revenue per token. The consensus treats this as a simple price war. The tiered threshold instead reflects cost-based segmentation. Anthropic cut prices deeply where serving costs are low and kept price discipline where the key-value cache, the memory that stores earlier tokens for reuse, is expensive to maintain. The precedent is OpenAI's GPT-4o mini launch in 2024, priced far below GPT-3.5 Turbo, after which rivals matched within months. The likelier outcome is that OpenAI and Google reprice their small tiers within weeks and adopt similar length-based tiers, as long as the Haiku 5.5 benchmark gains hold up in independent testing. The read is wrong if rivals hold their prices and Anthropic's share of small-model traffic does not rise, which would mean buyers are choosing on capability and integration rather than list price.
Part of a tracked trend
Frontier Model Price War
Frontier API vendors increasingly compete on strategic price-cutting rather than pure capability, repeatedly launching or repricing models below prevailing rates to grab share and compressing industry inference margins; expect recurring below-rival pricing moves.
Townsquare
Start a discussion in Townsquare.
More from this edition
- Nvidia releases Nemotron 3.5 Super VL with open weights, training data and reinforcement-learning environments
- OpenAI trains staff on spy tactics after federal prosecutors charge an engineer with hiding a recording device in an executive meeting room
- Reflection offers its 501-billion-parameter Beam model free on Ollama Cloud for a limited time, before releasing the weights
- Tencent open-sources Hy Memory 2.0, a 30-billion-parameter memory writer and a hub that collects coding-agent histories
- RealtimeBench shows a slower frontier model losing to a faster one once the game clock keeps running
- Pairwise judging of saturated-benchmark transcripts predicts held-out capability with 5.3 times less error, paper finds
- Anthropic folds Project Glasswing into a three-tier cyber program that relaxes classifier blocking for vetted security teams
Comments
0No comments yet.