Hey PAPAFAM ππΌ
Your hardest coding problem and your hundredth routine form submission probably should not use the same model.
This week's releases make that choice more interesting. OpenAI's GPT-6.1 Terra targets visual reasoning and computer use at lower API token prices than Sol. A separate Decisions API handles the tiny choices that can quietly make an app slow and expensive
π₯ The Big One
GPT-6.1 Terra brings cheaper computer use
OpenAI introduced GPT-6.1 Terra on 26 October. It combines visual reasoning, computer use and writing, with access through ChatGPT Work, Codex and the API as gpt-6.1-terra. The announcement lists Plus, Pro, Enterprise and Edu access for the product rollout.
The straightforward comparison: standard API input and output prices are half those of GPT-6.1 Sol and GPT-5.6 Terra. Terra costs $1 per million input tokens and $5 per million output tokens; cached input is $0.05 per million. That is API pricing, not a subscription price cut or a promise of twice as many completed tasks.
There is an important long-context caveat. Above 272,000 input tokens, the full request uses double the input and cache rates and 1.5 times the output rate. Check the model pricing details before budgeting a huge repository dump.
βFor the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice.β (OpenAI)
β‘ What shipped this week
1. OpenAI's Decisions API handles the small choices
The Decisions API became generally available on 22 October. You define a question and allowed answers; it returns a probability, a choice or a score. Think routing a support request, rather than asking a general chatbot to write an essay about where it belongs.
βYou only pay for input tokens: there are no output-token charges.β (OpenAI)
The announced base input rate is $0.10 per million tokens. It supports text, images and native JSON, with up to 400 questions per request. Start with one repeated decision and test the awkward examples. A confidence score is not a guarantee that the decision is right.
2. Vercel cut Agent session costs without cutting its rates
Vercel's 27 October update says dashboard and Slack Agent sessions now cost 67% less on average. The changes are live: better context management, prompt-cache reuse and clearer instructions for common workflows.
βThese improvements are live and require no action.β (Vercel)
Billing rates are unchanged. This is Vercel's reported average reduction in work consumed per session, not a discount on every token or a guaranteed saving for your workload. The lesson for builders: inspect what your agent repeatedly reads before assuming the model is the whole cost problem.
3. Google Docs can now edit native Markdown files
Google's 26 October update lets you upload an .md file to Drive and edit it directly in Docs without converting it into another file type. Comments, collaborative editing and version history come with it. Google says anyone with a Google or Workspace account can start.
βYou can edit without changing file typesβ (Google)
This is useful when your agent works in Markdown but your teammate reviews in Docs. Keep one canonical brief and check the exported file before feeding it back into a repository. Native editing does not mean automatic Git synchronization.
4. Agent Skills gets a vendor-neutral home
On 22 October, Anthropic announced it is donating Agent Skills to the Agentic AI Foundation under the Linux Foundation. A skill is a folder containing instructions, plus optional scripts and references for a specific job.
βYou can continue to use the standard or contribute to it without joining the AAIF.β (Anthropic)
The practical opportunity is preserving your team's working knowledge outside one chat or vendor. Turn a repeatable workflow into a small, reviewable skill. Read any downloaded scripts before running them: a shared format is not a security endorsement.
π§° Worth your time
-
Heal: Runs your app in a sandbox and reports reproducible defects with failing tests. A discovery from this week's launch notes, not a verified new-release date. The current site offers a free trial; do not assume unlimited free beta access.
-
MiniMax M3.1 Flash: Another low-cost agent model to evaluate, at $0.15 input and $0.60 output per million tokens. Discovered in this week's notes, but the vendor launch was 19 October. Its announcement says model weights will come later.
-
ElevenLabs on Vercel AI Gateway: Speech, transcription and realtime voice models through one gateway key. Vercel's 27 October announcement offers 50% off ElevenLabs models on that gateway through 9 November, not a permanent provider-wide price cut.
This week's challenge: pick one repetitive step in your app or agent. Write down what a correct result looks like, test a cheaper route, and keep the difficult examples in the test set. Measure the whole job, not just the first response.
Which task would you move to a cheaper model first, and what mistake would make you switch it back? Hit reply and tell me.
I read every single one.
Talk soon PAPAFAM,
Sonny ππΌ
ππ½ Don't forget to follow me across socials!
Responses