OpenAI ships GPT-6.1 Terra to Bedrock and Copilot on day one, Anthropic files for an IPO, Cloudflare discloses a GLM-5.3 cross-request leak
The thread today: a cheaper model tier lands on every channel at once, and the real risk moves to how models are served. GPT-6.1 Terra appeared on OpenAI, AWS and GitHub at half of Sol's price, while Cloudflare's incident shows one configuration mismatch in self-hosted inference can leak data across customers. Choosing a model keeps getting easier; the engineering discipline of running one matters more.
Released on October 26 as gpt-6.1-terra on Responses and Chat Completions. OpenAI describes it as fast and efficient, for visual reasoning, precise computer use and clear, natural writing.
Standard pricing is $1 input, $0.05 cached input and $5 output per million tokens; on the same pricing page gpt-6.1-sol is $2 and $10. Batch and Flex are 50% lower, and Fast mode is 2x Standard.
Context window of 1,050,000 tokens, up to 128,000 output tokens, and a June 1, 2026 knowledge cutoff.
Prompts over 272K input tokens are billed at 2x input and cache rates and 1.5x output for the whole request. Cache writes cost $1.25 per million for 30-minute retention and $2.50 for 12-hour retention.
Builder's takeNext to the 12-hour prompt caching in the 10-21 issue, OpenAI keeps slicing its price tiers finer. I would move part of PandaClaws' rewriting and caption calls, which do not need the strongest reasoning, to Terra and compare quality and unit cost against Sol on my own samples. Keep long documents under 272K tokens, or the whole request is billed at a premium. My post on evaluating model releases covers how to run that test.
AWS announced on October 26 that GPT-6.1 Terra is generally available on Bedrock, with up to 1 million input tokens of context and text and image input.
On OpenAI's Agents' Last Exam evaluation it scored 15 percentage points above GPT-5.6 Terra, on par with GPT-6 Astra, and finished tasks more than 3x faster than GPT-6 Astra and GPT-6.1 Sol.
Input and output prices are at least 50% lower than GPT-6.1 Sol and GPT-5.6 Terra.
It can be called through Bedrock APIs, and ChatGPT Work and Codex can be configured to use the model on Bedrock.
Builder's takeWith OpenAI's new model on AWS on day one, and Google Cloud announced on 10-21, which model you use and which cloud you buy it from are now separate decisions. Teams billed through AWS can try Terra in Bedrock directly, but note the Bedrock page says 1 million input tokens versus 1.05 million on OpenAI's own API, so check limits and prices on the channel you actually call.
Generally available to Copilot Pro, Pro+, Max, Business and Enterprise users, positioned for everyday agentic coding and terminal work at lower cost than GPT-6.1 Sol. In early testing it solved noticeably more terminal and C# coding tasks than GPT-5.6 Terra.
Billed at provider list pricing under usage-based billing.
Selectable in the model picker in VS Code, Visual Studio, Copilot CLI, the cloud agent, the Copilot app, github.com, GitHub Mobile, JetBrains IDEs, Xcode and Eclipse, with a gradual rollout.
Business and Enterprise admins manage it through the model policy. Under default model enablement, new models are turned on automatically unless an admin has turned off the global default or disabled this model.
Builder's takeThis is the first new model since the global default policy from 10-23 took effect: in organizations that never configured it, Terra will simply appear in everyone's picker. I would watch a week of the per-client usage report from 10-25, compare unit cost and merge rate for routine tasks on Sol versus Terra, and only then decide whether to make it the team default.
On October 26, Anthropic, PBC filed a Form S-1 registration statement with the U.S. SEC for a proposed initial public offering of its Class A common stock.
It intends to list on the Nasdaq Global Select Market under the symbol "ANT".
The number of shares and the price range have not been determined; the registration statement is not yet effective, and no shares may be sold before it is.
The preliminary prospectus will be available from Goldman Sachs and Morgan Stanley.
Builder's takeI am not going to talk valuation. The useful part for people building products: once the prospectus is public, it will be the first systematic description of the business and its risks written by Anthropic itself. I will read what it says about the API business, compute supply and usage policy to judge Claude's long-term stability as a dependency. Until then, do not change any technical choice because of the headline.
Between August 27 and September 9, 155 responses from GLM-5.3 on Workers AI contained data from other customers' earlier requests, all to prompts longer than 681,280 tokens. Cloudflare found no evidence of deliberate exploitation.
Gadi Evron of Knostic reported it via HackerOne on September 8. Cloudflare took the model offline within 40 minutes of declaring an incident, restored it the same day, and published on October 26 in coordination with the researcher and the SGLang maintainers.
Two causes: the model launched with a 1,048,576-token context while the prefill KV pool held only 681,280 tokens, and tests checked only HTTP status codes; and in disaggregated prefill/decode mode SGLang swapped a rejected request for a one-token placeholder, so decode read stale cache pages and the client got a 200 instead of a 400.
After the fix the pool holds about 2.04 million tokens. SGLang patched the bug in v0.5.20 with follow-ups in v0.5.21 and v0.5.22, and Cloudflare recommends all operators upgrade to v0.5.22 or later.
Builder's takeThis is the kind of risk self-hosted open models miss most easily: no error, just data quietly crossing tenants. If you run SGLang, upgrade to 0.5.22 first. Before launch, read the KV pool size from the running engine and confirm it is at least the advertised context, and make long-context tests check the output itself, with canary strings, not just status codes. For multi-tenant file Q&A like AI Cloud Drive, this goes on my release checklist.
From October 26, new projects using the Generate Content API (generateContent, streamGenerateContent and batchGenerateContent) on the paid tier store requests by default (store=true), matching the Interactions API.
Existing projects keep their current setting.
Opt out with store=false on individual requests, or turn off Generate Content API logging in the Settings panel of the Logs and Datasets dashboard.
Builder's takeThis hits new projects, so the likely victims are freshly created test or client projects. I would pass store=false explicitly in the call layer rather than rely on a console setting. For a product like AI Cloud Drive that handles people's private files, whatever the privacy notice promises has to be switched off in code to match.
View and edit .md files directly in Google Docs without converting them; a .md badge in the header marks a native Markdown file.
People and AI agents can co-edit the same file, with line-level comments, suggestions and a version history covering both human and agent edits.
Drive renders Markdown previews with tables and links, and can toggle to raw text.
Anyone with a Google or Workspace account can start by uploading an .md file to Drive.
Builder's takeMore and more agent-written plans and docs are Markdown, and reviewing them meant either reading diffs in Git or converting to Docs and back. Now non-technical colleagues can comment on the .md file itself in Docs. I will try this for PandaClaws long-form drafts, but the announcement says nothing about syncing with a Git repo, so getting files into the repo is still on you.
Adds 25 languages, including 11 Indic languages plus Catalan, Malay and Welsh, for 50 languages across 60+ locales in total.
Across internal and public benchmarks such as GigaSpeech and VoxPopuli, token error rate improves by up to 25% and entity error rate by up to 37%, most on financial-domain entities.
Up to 2x faster than 2607, especially for 5-10 second clips at high concurrency, with fewer hallucinations on silent or empty audio.
Multilingual mode is the default: omit locale hints when one file mixes languages; hints are for genuinely single-language audio.
Builder's takeFor voice interviews, entity error rate is the number to watch: get a company or technology name wrong and every downstream score drifts. Mixed Chinese and English is common in AI Interview recordings, so I would drop the locale hint as advised, measure proper-noun errors on my own real recordings, and compare against the current setup rather than looking only at overall error rate.
Bedrock supports implicit caching, where it chooses where the reusable prefix ends, and explicit caching marked with prompt_cache_breakpoint. AWS recommends both: explicit for tools and instructions, implicit for the growing history.
The reusable prefix must reach 1,024 tokens. In one test, a roughly 4,000-token tools-and-instructions prefix wrote 4,120 tokens on the first request and read those 4,120 tokens on later ones.
Caching only hits on identical text: serialize tools and instructions the same way every time, keep MCP-delivered tools in a stable order, and put variables such as the date, user name or request ID after the cached content.
store=False only stops response storage on the service side; it does not disable prompt caching.
Builder's takeThese rules apply well beyond Bedrock: putting a timestamp at the top of the system prompt is the most common cache-buster I see. In my post on a month of real RAG bills, controlling context length paid back the most, and agents are the same: make tool definitions and instructions a fixed prefix, watch the cache hit rate, and only then think about switching models.
Researched and drafted with AI assistance; editorial standards and views set by Darius.
Every Monday, the week in AI in one email
Every Monday morning: the 6 AI stories that mattered most in the past seven days, with my take, in your inbox.
Your email is only used to send the weekly and is never shared. Every email has a one-click unsubscribe link. See the latest issue →