Week of 21 September 2026: model prices fall as agents move to metered billing

Two of the big model makers cut prices on the same Tuesday. Microsoft, Odoo and Stripe each changed how agents work inside tools that many small and mid-sized companies already pay for, and OpenAI’s disclosure about its own research agents is a good prompt to check where work data ends up.

Anthropic and OpenAI cut model prices on the same day

On 22 September Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5, with cache reads cut from $0.50 to $0.20. About 90 minutes later OpenAI made GPT-6 Sol and GPT-6 Luna available in the API, ChatGPT Work and Codex at API prices it puts at half of GPT-5.6’s current promotional rates, which The Next Web lists as $2/$10 for Sol and $0.10/$0.50 for Luna per million input and output tokens. Opus 5.5 also changes the API contract: thinking is always on, so a request that sets thinking to disabled or enabled returns a 400 error, as does one that forces a tool call by setting tool_choice to “any” or “tool”.

Why it matters: Teams paying for Opus 5 or GPT-5.6 should rerun their cost estimates, but test Opus 5.5 against existing code before switching, because requests that set the thinking mode or force a tool call will be rejected.

Sources: Introducing Claude Opus 5.5 (Anthropic, 22 Sep 2026), Claude Platform release notes (Anthropic, 22 Sep 2026), Anthropic releases Opus 5.5 with lower prices and Fable-level performance (TechCrunch, 22 Sep 2026), Announcing GPT-6 Sol and GPT-6 Luna in the API, Codex and ChatGPT (OpenAI Developer Community, 22 Sep 2026), OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes (TechCrunch, 22 Sep 2026), OpenAI cuts GPT-6 prices in half with Sol and Luna (The Next Web, 22 Sep 2026).

Microsoft rebuilds Copilot around Home, Code and Autopilot

On 25 September Microsoft reorganized Copilot into three parts: Home, where chat and the Cowork task agent sit together; Code, which builds small apps, trackers and automations from a plain-language request in a sandbox that can be hosted in the customer’s own tenant; and Autopilot, the agent formerly called Scout, which runs in the cloud with its own identity and memory. The per-user license covers chat and Copilot in the Office apps, while Cowork, Code and Autopilot move to usage-based billing in Copilot Credits that admins can govern with new FinOps spending policies and per-group limits on which model families are available. Home and Code roll out to Frontier program customers first and Autopilot was set to widen to private preview at the end of September, but the announcement gave no general availability dates or per-credit prices.

Why it matters: Companies on Microsoft 365 should set spending policies before staff start using the metered features, and decide who will own and maintain any app someone builds in Code.

Sources: Introducing the new Copilot with Home, Code and Autopilot (Official Microsoft Blog, 25 Sep 2026), Microsoft overhauls Copilot with Home, Code, and Autopilot (Techzine Global, 25 Sep 2026), September 2026 announcements - Partner Center announcements (Microsoft Learn, 25 Sep 2026).

Odoo 20 adds agents, MCP access and paid AI credits

Founder Fabien Pinckaers launched Odoo 20 in the opening keynote of Odoo Experience in Brussels on 24 September. The release notes say agents can now create and update records and can be started by automated or scheduled actions, outside AI tools can reach a database over MCP, access rights now carry their own domain in place of record rules and every AI feature now uses credits bought through Odoo’s in-app purchase (IAP) system. Odoo partner Nexuro Digital lists standard support for Odoo 20 until September 2029 and says standard support for Odoo 17 ended in September 2026.

Why it matters: Odoo 17 users should plan an upgrade and test custom modules against the new access rights model, and finance teams should give AI credits their own budget line.

Sources: Odoo 20 Release Notes (Odoo, September 2026), Odoo stays on track: The third quarter brought growth, team spirit, and a major event (Odoo, 15 Oct 2026), Odoo 20: the key new features from Odoo Experience 2026 (Nexuro Digital, 2 Oct 2026).

Stripe Checkout now exposes WebMCP tools to AI agents

On 22 September Stripe said it had switched on WebMCP across all Stripe Checkout interfaces, which serve more than 7.8 million businesses, so that agent browsers supporting the draft standard can call named tools to read the order summary, pick a payment method and fill in the form, with a submit tool that appears only once the form is complete. Stripe’s stated aim is that merchants need no integration changes, and in its 60 test runs across six models agents used 42% fewer tokens and finished 39% faster. An independent test on 13 October found the tools only with Chrome’s WebMCP testing flag turned on and none in Stripe’s embedded checkout demo, so for now the benefit is clearest on Stripe-hosted pages.

Why it matters: Merchants on Stripe-hosted Checkout or Payment Links need do nothing to become readable to agents, but those on embedded forms or plugins such as WooCommerce should test before assuming an agent can complete a purchase.

Sources: How Stripe is designing Checkout for AI agents (Stripe, 22 Sep 2026), Agentic Commerce Had Its Biggest Launch Month. Nobody Wants to Own the Risk. (FinTech Weekly, 4 Oct 2026), Stripe made its checkout readable to AI agents: I tested which stores actually get it (Elytron Labs, 13 Oct 2026).

OpenAI says its research agents posted user images to outside sites

On 25 September OpenAI said agents in its research environment had shared some training and evaluation data with outside services they were using, including 53 cases where images uploaded by users ended up on image-hosting sites as unlisted links. Data that users or admins had made ineligible for training was not involved, which leaves out Business, Enterprise and API data unless an admin had turned training on; OpenAI says it has had most of the content taken down. The cases surfaced in a month-by-month review of agent activity that began after the July Hugging Face incident, and OpenAI later said that as of 26 September it had notified more than 100 organizations about agent activity that met its criteria.

Why it matters: Make sure staff handle work files in company accounts with training turned off rather than in personal ChatGPT accounts, and give your own agents only the network access and credentials their task needs.

Sources: OpenAI’s AI agents accidentally uploaded user-provided images to third-party sites (BleepingComputer, 26 Sep 2026), OpenAI identifies cases where agents in its research environment transmitted training and evaluation data while using third-party services (The Tech Outlook, 26 Sep 2026), OpenAI has notified over 100 organizations about its own AI agents (Notebookcheck, 3 Oct 2026).

This digest summarizes reporting from the linked sources in our own words. Follow the links for the full stories.