The cheap tier of the model market got crowded this month. Anthropic released Claude Haiku 5.5(opens in a new tab) on October 7 at $0.10 per million input tokens and $0.50 per million output tokens, on AWS, Google Cloud and Azure. On October 19, MiniMax launched M3.1 Flash(opens in a new tab) at $0.15 in and $0.60 out, and Inception launched Mercury 3(opens in a new tab), a diffusion model, at $0.20 in and $0.70 out, half that during a launch promotion.
The list prices sit within a factor of two. For a workload of one billion input tokens and 200 million output tokens a month, that is roughly $200 on Haiku 5.5, $270 on M3.1 Flash and $340 on Mercury 3 at standard rates. At this tier, the token bill is rarely the number that decides the business case. Staff time spent waiting on the answer, and fixing wrong answers, is.
The real differences are elsewhere. Inception reports 1,068 tokens per second on NVIDIA GPUs and claims four to eight times the speed of Haiku 5.5, with a one million token context window. MiniMax reports 170 tokens per second at the median. Haiku 5.5 is the only one of the three sold through all three hyperscalers, but its price rises fivefold, to $0.50 in and $2.50 out, once a prompt passes 100,000 tokens. A retrieval or document workload with long prompts can cost several times what the headline rate suggests.
Treat the benchmarks with care. Inception's launch table puts Mercury 3 ahead of Haiku 5.5 on 10 of 12 benchmarks, but the scores are its own, and Haiku leads on long-context recall and one of the two coding tests. MiniMax says its weights will come later, with no date, and a buyer subject to data residency or China-origin procurement rules will need to settle where it runs before testing it.
The practical move is to stop choosing one small model by list price. Run your own top three high-volume tasks through each, and measure cost per completed task, latency and error rate at your real prompt length. Keep contracts for this tier short and model-agnostic, because a cheaper or faster option has arrived roughly every week this month.
Action items
Three small models launched in 12 days at list prices within 2x of each other. Price no longer separates them; speed, prompt length and origin do.
For the CFO
At this tier the token bill is small. Measure cost per completed task, including staff time spent fixing errors.
For the CTO
Benchmark your top three high-volume tasks on each model at real prompt lengths, and keep routing model-agnostic.
For the CISO and General Counsel. Settle data residency and China-origin procurement rules before anyone tests MiniMax, and cap what agents may disclose.
For the CAIO
Keep small-model contracts short. A cheaper or faster option has arrived nearly every week this month.
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
- InceptionIntroducing Mercury 3(opens in a new tab)
A diffusion model at 1,068 tokens per second, $0.20 in and $0.70 out, with a 1M token context. Available via OpenRouter, Baseten and Vercel. Speed claims are the vendor's own, so test before routing live traffic.
- MiniMaxMiniMax M3.1 Flash(opens in a new tab)
Priced at $0.15 in and $0.60 out, with MiniMax claiming agent tasks cost 70% less than its prior model. Weights are promised later with no date, so self-hosting for residency is not an option yet.
- Mistral AIMistral releases Timestral, its first time-series foundation models(opens in a new tab)
Four open models, from 5M to 209M parameters, under a modified MIT license, forecasting on a CPU in as little as 10 ms. Treasury, supply chain and plant teams can test zero-shot forecasting on-prem without a GPU budget.
- Hugging Face PapersOfficeTown: Generative Organizations for Evolving Agent Intelligence(opens in a new tab)
On 406 simulated office tasks across 139 job titles, the best model passed 74.4% on the first try and Claude Opus 5.5 passed 50.7%. Plan for a human to finish a quarter or more of delegated knowledge work.
- FortuneMeta's AI agent gave a Facebook Marketplace buyer a seller's home address(opens in a new tab)
With 'always allow' permissions, Meta's Muse agent shared a user's home address and appeared to offer a price $100 below his set minimum. Any agent you deploy with standing permissions needs hard limits on data it can disclose and terms it can accept.