Anthropic released Claude Haiku 5.5 on October 7 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, according to its launch page(opens in a new tab). Haiku 4.5 costs $1 and $5. Anthropic says the new model costs around 75% less to run on average. List prices are 90% lower on requests up to 100,000 tokens and only 50% lower on longer ones, where the rate rises to $0.50 and $2.50.
A day earlier, OpenAI opened a public beta of a Decisions API that runs only on GPT-6 Luna and charges $0.10 per million input tokens(opens in a new tab), with no charge at all for output or cache. It returns a typed answer, not prose: a probability that a statement is true, a pick from a fixed list with a confidence score, or a score on a set range. OpenAI says it decides up to 10x faster than the same model through its Responses API.
Put those together and the price of a routine judgment (is this invoice a duplicate, which queue does this ticket go to, does this clause match our standard) now starts at $0.10 per million input tokens at both vendors. For Anthropic customers that is up to 90% below the previous small model. In many pilots, that is where most of the token volume sits. It is also work most teams still run on a mid-tier or frontier model because that is what the first prototype used.
The fine print matters for the business case. Anthropic's own footnote says its 75% average already allows for a new tokenizer that uses slightly more tokens per task, so the 90% list-price cut overstates the per-task saving. Long-document work sits above the 100,000 token line and only gets half off. And on Anthropic's benchmark table(opens in a new tab), Sonnet 5.5 beats Haiku 5.5 on every listed test, with Anthropic itself pointing complex agentic coding to Sonnet and Opus. The saving is real for short, repetitive, checkable tasks. It is not a reason to downgrade everything.
The practical move is a routing review. List your top ten AI workloads by monthly token volume, mark which ones are short classification or extraction, and rerun a sample of each on the cheap tier against the same acceptance test you used to approve the pilot. Where quality holds, move the volume and book the difference. Where it drops, you now have a measured reason to stay on the larger model, which is a better position at renewal than an assumption.
Action items
The floor price of a routine AI judgment dropped to $0.10 per million input tokens this week. In many pilots, that is where most of the token volume sits.
For the CFO
Ask for the top ten AI workloads by token volume and what each one would cost on the new small-model tiers.
For the CAIO
Rerun short classification and extraction tasks on the cheap tier against the original acceptance test, and move the ones that pass.
For procurement
Price per task, not per token. A new tokenizer and a long-prompt surcharge both shrink the headline discount.
Researched and drafted by an automated workflow, then reviewed and edited by a human editor before publication. Every source is linked. See how we use AI here.
Also worth knowing
- AnthropicIntroducing Claude Haiku 5.5(opens in a new tab)
Haiku 5.5 lists at $0.10 input and $0.50 output per million tokens under 100,000 tokens, against $1 and $5 for Haiku 4.5. Rerun your short, high-volume tasks on it before your next spend commitment.
- OpenAI Developer CommunityDecisions API is now available in public beta(opens in a new tab)
OpenAI bills only input, at $0.10 per million tokens, for typed probability, choice and score answers on GPT-6 Luna. Routing and triage steps inside agent workflows are the obvious candidates to price against it.
- StocktwitsMSFT, ORCL, NVDA, QQQ In Focus: OpenAI's Annualized Revenue Reportedly Trails Prior Expectations By $20B(opens in a new tab)
The FT put OpenAI near $50 billion annualized, about $20 billion under a widely reported figure. Anthropic counts cloud partner sales and OpenAI does not, so do not compare vendor scale on headline run rates.