Issue #3 ·
Faster, cheaper models, a COBOL milestone, and money for checking AI code
Welcome back. Since Monday the pattern is clear: models are getting faster and cheaper for the many small steps an agent takes, and money is flowing to the tools that check what agents produce. Every item below was checked against the original source, and the numbers are exactly as the source gives them.
Top stories
1 · Models
Inception's Mercury 3 runs at 1,068 tokens per second
Inception launched Mercury 3 on October 19, a diffusion language model that writes text in parallel passes instead of one token at a time. The company says it reaches 1,068 tokens per second on NVIDIA GPUs, has a 1M-token context window (up from 260K for Mercury 2.5) and beats GPT-6 Luna and Claude Haiku 5.5 on a range of benchmarks while running 4 to 8 times faster. List price is $0.20 per million input tokens and $0.70 per million output tokens, with 50% off at launch ($0.10 / $0.35). It is OpenAI API compatible and available on Inception's API, OpenRouter, Baseten and Vercel.
Why it matters: An agent waits for the model at every step, so speed adds up fast. Because it uses the OpenAI format, testing it on one slow step costs you an afternoon. Budget at the list price, not the launch discount.
Source: Inception blog
2 · Dev tools
OpenAI's newest models can now migrate COBOL
On October 19 OpenAI published results for older, less documented languages such as COBOL, Fortran and Visual Basic. In its COBOL migration test, where the rewritten program must produce byte-for-byte the same output, GPT-6.1 Sol scored 86.6% at low reasoning effort, against 5.4% for GPT-5.6 Sol, at 61% of the cost. At higher effort it reached 97.9%. On MultiLangBench, which asks a model to build a SQLite-compatible database in 20 languages, GPT-6 Astra averaged 87.4%, up from 26.8%. OpenAI notes that production migrations still need expert review and testing.
Why it matters: Old systems that nobody dares touch are now a real market for migration work. The method is the lesson too: decide what “correct” means as an exact output test before you let an agent rewrite anything.
Source: OpenAI
3 · Models
Cloudflare's Clairic pulls typed values out of messy text
Cloudflare released a preview of Clairic on Workers AI, a model that does not write text. You send a document and a list of typed fields (a number, a date, a list, an estimate) and it returns one value per field, with a confidence score and the chance the value is missing, or null if it is not there. It costs $0.10 per million input tokens and has no output charge. In Cloudflare's damaged-order demo, a Clairic, Clef and GLM-5.3 pipeline finished in 19.3 seconds against 61.7 seconds for GLM-5.3 alone and cost 49% less. Cloudflare says multi-step arithmetic is still weak in this early version and plans to release open weights soon.
Why it matters: Invoices, order emails and refund requests are mostly “find the number” jobs. A typed extractor with a confidence score tells you which ones are safe to automate and which need a human.
Source: Cloudflare blog · Changelog
4 · Models
MiniMax M3.1 Flash: agent work for $0.15 per million input tokens
MiniMax rolled out M3.1 Flash on October 19 at $0.15 per million input tokens, $0.60 per million output tokens and $0.015 for cached input. MiniMax says usage cost on agentic tasks is 70% lower than its predecessor and reports 73.7% on the DeepSWE v1.1 coding benchmark. The anonymous “Space Bunny” model that processed more than 200 trillion tokens was an early version. It is available through the MiniMax API, M Plan and MiniMax Code, and MiniMax says the weights will be released later.
Why it matters: For high-volume steps like tagging, routing and first drafts, a model at this price changes the maths of what you can automate. Run your own test set before switching, and check the license when the weights arrive.
Source: MiniMax blog
5 · Funding
Greptile raises $50M to check code before it merges
Greptile, which reviews and tests code changes, raised a $50M Series B led by Menlo Ventures at a $750M valuation, with Jump Trading, Wing, Operator Collective, Benchmark, YC, Browder Capital and Initialized joining. It says it reviews and tests more than 1% of all pull requests opened on GitHub every day, for teams including Nvidia, Coinbase, Decagon and Vanta. Its stated mission is to automate code validation so that even teams building mission-critical software can auto-merge complex changes.
Why it matters: As agents write more of the code, the scarce part is checking it. Whatever agents do in your business, a separate checker that is not the same agent is worth paying for.
Source: Greptile blog
Open source & open weights
- •Ornith-1.5-122B is an open-weight mixture-of-experts model under the MIT license. Its model card says it serves on a single 80GB GPU, handles 262,144 tokens of context and scores 84.5 on SWE-bench Verified in the team's own tests. Hugging Face
- •Meta published the technical report for Muse Glimmer 30B, an on-device agent model that takes text and images and is quantized to about 4-bit for single-GPU use. The weights have been out under Apache 2.0 since August; the report is the new part. Meta research
- •Headroom compresses tool outputs, logs, files and RAG chunks on your own machine before they reach the model. It claims 20% fewer tokens for coding agents and 60–95% fewer for JSON (Apache-2.0, 76,819 stars). GitHub
Quick hits
- •ChatGPT is rolling out locked chats: hide chosen conversations from history and search behind a 4-digit PIN or device login, on Free, Go, Plus and Pro plus Business, Enterprise and Edu. Retention settings do not change. Release notes
- •OpenAI acquired Orchestra Bio, whose agents track drug-development programs, and is folding it into Rosalind, its life-sciences models and tools. OpenAI
- •Anthropic had an internal Claude research model design protein binders against scorpion neurotoxins over 64 hours. Labs tested 211 designs and 17 met the binding criteria. Anthropic stresses this is not an antivenom yet and is releasing all 211 designs and their data. Anthropic
- •Nace AI's Drex Max is a decision model that reads text and up to 8 images per request, at $0.06 per million input tokens with no charge for output. Drex docs
- •ChatGPT Slides lets a team edit and comment on the same deck and tag ChatGPT in a comment to make changes, then export to PowerPoint or Google Slides. Pro, Business and Enterprise plans. ChatGPT Slides
Repo of the issue
DeepSeek open-sourced the inference framework behind its production services. This release does offline inference for DeepSeek-V4.1-Flash on NVIDIA SM100 GPUs and Huawei Ascend 950 chips, with structured outputs and deterministic sampling. A KV cache system, prefill/decode disaggregation and load balancing are not included, so it is not a full serving stack yet. It was originally built mainly on vLLM. On October 21 it had 14,013 stars and it is Apache-2.0 licensed.
For founders
- 1Match the model to the step. Fast, cheap models for the many small steps, a strong model only where judgment matters, and measure each step on your own data.
- 2Don't ask a chatbot to do your bookkeeping. Pull numbers and dates out with a typed extractor, do the maths in code, and use an LLM only for the words.
- 3Pay for checking, not just for doing. Greptile's raise and OpenAI's byte-for-byte tests point the same way: define what correct looks like before an agent ships work.
From Agentvendo
Running a Shopify store? The Agentvendo packs are ready-made AI agents for store owners: daily P&L, support replies, abandoned-cart emails and more, plus the new Approval Queue & Kill Switch so every refund, price change or customer email waits for your OK. Starter is $29 one-time and Complete (all 9 templates) is $79.
See the packs →