Daily AI brief · September 22, 2026
Claude Opus 5.5 Ships, 40% Cheaper Than Opus 5
Anthropic's Opus 5.5 matches Fable 5.1 on most work at lower cost, Amazon blocks Meta's shopping agent, and British Columbia sues OpenAI over unreported ChatGPT threats.
Anthropic releases Claude Opus 5.5 (Anthropic, 9to5Mac) Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5 on typical workloads. API prices drop 20% to $4 per million input tokens and $20 per million output, cache reads fall 60% to $0.20, and output is about 30% faster. It's available on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry, with Sonnet 5.5 and Haiku 5.5 due in the coming weeks. METR and the US Center for AI Standards and Innovation tested it before release. What it means for SMBs / MSPs: If you run Opus 5 through the API, switching the model ID is the quickest cost cut you'll get this quarter, but rerun your own test prompts first.
Amazon blocks Meta's Muse agent from shopping on Amazon.com (GeekWire, Forbes) Amazon cut off Meta's new Muse personal agent after Meta declined to exclude the site. Amazon says Muse doesn't identify itself when it browses and appears to store customer credentials. Meta says Muse can't see passwords or payment methods. Amazon has also sued Perplexity and moved to block shopping agents from Google and OpenAI. What it means for SMBs / MSPs: Agents that shop for your clients' customers will be allowed on some sites and blocked on others, so don't build a buying workflow that depends on one marketplace tolerating it.
British Columbia sues OpenAI and Sam Altman over the Tumbler Ridge shooting (Al Jazeera, Tom's Hardware) The province filed in federal court in San Francisco, alleging OpenAI flagged the shooter's ChatGPT account in June 2025, that reviewers recommended a police referral, and that leadership declined because the case didn't meet its threshold for a "credible and imminent" threat. Eight people died in the February 10 attack. The province is seeking recovery costs and changes to how OpenAI handles violent conversations. What it means for SMBs / MSPs: Any client putting a chatbot in front of the public needs a written escalation path for threatening conversations before launch, not after an incident.
Xiaomi open-sources MiMo-V2.6 under the MIT license (SiliconANGLE, eWeek) Xiaomi released the MiMo-V2.6 family, including a 1.02-trillion-parameter Pro model with 42 billion active parameters that takes text, image, video and audio. It also published more than 7,000 reinforcement-learning environments and its training framework. The MIT license allows commercial use. What it means for SMBs / MSPs: Permissively licensed frontier-scale models keep arriving, though the Pro model needs data-center hardware; for most clients, the smaller siblings are the ones to test.
An Opus 5.5 swap test. This is for champions and MSPs with any Claude API workload, such as document summaries, ticket triage or internal chat. It takes about 45 minutes. First step: pick ten real prompts from last week, run them on Opus 5 and on Opus 5.5 side by side, and compare quality, speed and the token cost shown in your console. If the answers hold up, change the model ID in one workflow and watch it for a week.
If you pay for Claude through the API, ask your developer where the model ID is set. Anthropic says Opus 5.5 costs 40% less to run than Opus 5 on typical workloads, with API prices down 20%. Changing the model ID may be your quickest cost cut this quarter. Rerun ten real prompts on both versions first, then switch one workflow and watch it for a week.
*For MSPs:* A model-refresh check that finds hard-coded model IDs and reports cost per task before and after is quick to deliver.
Taking "matches Fable 5.1 on most work" as "matches it on your work." "Most" is a vendor's average. Test the tasks your client depends on, especially long documents and multi-step agent runs, before you retire the more expensive model.
Frontier capability is getting cheaper faster than it's getting better. Each new tier now lands near the previous top model's quality at a lower price. For small businesses, the payoff is that last year's "too expensive to automate" tasks are worth revisiting every few months, and the teams that re-test regularly will capture the savings first.
Compiled on September 27, 2026 from reporting published on or before September 22, 2026.