Dan O’Brien

Any enterprise working on @AnthropicAI Claude Opus for agent work gets more of the work done for less money by moving to yesterday's brand-new Opus 5.5, and in our testing the gap is...pretty large. Opus 5.5 gets more work done for more people, at a noticeably lower price than Anthropic offered before, as measured by @Signal_65 PINNACLE. ➡️ 2.3x fewer weighted errors than Claude Opus 5, and 100% of the multi-step jobs finished end to end against 95% ➡️ 34% less per correct task (!!), $0.40 against $0.60, on a list price 20% below Opus 5 with cache reads at 5% of input rather than 10% ➡️ Fabrication in retrieval answers down from 7.6% to 2.8% ➡️ It also passes Claude Fable 5.1 with 1.3x fewer weighted errors at 63% less per correct task Full comparison on the Signal65 PINNACLE results views at pinnacle.signal65.com

💸💸Closed Model AI pricing is broken, and the vendors proved it with their own product lines 🤖🤖 Hosted AI crossed a threshold this month, in intelligence and in stability at once. In our testing the newest models from @OpenAI @Google and @AnthropicAI finish 98% or more of the multi-step enterprise jobs they start and rarely invent an answer, and they do it for roughly the same dollar per finished job the previous tier charged. @Signal_65 PINNACLE is our independent benchmark that scores AI models on real multi-step enterprise work and prices every correct task on the full API bill, or on the rented GPU node an open model runs on. The three snapshots connect each lab's models from the cheapest correct task upward. OpenAI put the new generation at the old price, and in our testing the ladder goes straight up between GPT-5.6 and GPT-6. ➡️ GPT-6 Astra at medium effort costs the same per correct task as GPT-5.6 Sol, about 94 cents, with 2.3x fewer weighted errors. A list price 2.5x higher was cancelled by 2.9x fewer input tokens per finished job. ➡️ Maximum effort on Astra buys another 1.8x fewer errors for 1.6x the cost and the top score on the board. The same dial on Sol bought nothing, so reasoning effort only pays on the new model. ➡️ GPT-5.6 Luna clears the open-weight baseline for 7 cents a task, the cheapest correct task from any hosted model we have tested. The Anthropic ladder has the new generation at the top and a broken rung at the bottom in our testing. ➡️ Claude Haiku 4.5 lists at half the price of Claude Sonnet 5 and costs more per correct task, $1.09 against $0.62, because it finished 29% of the jobs and every failed attempt is billed. A router that sends easy work to the small model sends it to the dearest rung here. ➡️ Sonnet 5 is the value rung of the line. Claude Opus 5 costs 2.2x Sonnet for 1.2x fewer weighted errors. ➡️ Claude Fable 5.1 is the Anthropic entry in the new generation and the one that did not arrive at the old price, 1.7x fewer errors than Opus for 1.8x the cost and the most expensive correct task from any hosted model on the board at $2.46. Google shows the whole curve, four open-weight Gemma 4 builds priced on a rented B300 node and two settings of hosted Gemini 3.8 Flash, and in our testing the ladder turns in the middle. ➡️ The Gemma builds get cheaper per correct task as they get better, because job completion climbs from 17% to 54% faster than the token bill grows. Below the turn the better model is also the cheaper one and there is no trade-off to make. ➡️ Gemini 3.8 Flash at medium effort makes 11x fewer weighted errors than the best Gemma for 1.5x the cost, 54 cents against 35. Inside the Google catalog the hosted Flash beats running the open weights yourself at the same money. ➡️ The high effort setting adds 16% fewer weighted errors for 20% more. Above the turn you are paying for reasoning, not completion. The score and the price of a correct task for every model we have tested, hosted and open, are live on the PINNACLE results views at pinnacle.signal65.com