AI & Tech Brief ⚡

📌 Navigate

📊 Exec Summary

Vendors priced model output by its shape this week, from typed decisions and on-device vectors to a speed tier and open weights, while 2 preprints put AI agents and design models into enzyme discovery.

Seven things moved in AI/tech this week:

• Anthropic priced Claude Haiku 5.5 at GPT-6 Luna's base price. Prompts under 100K tokens cost $0.10 input, $0.50 output and $0.01 cache reads per 1M, Anthropic reports 72.4% on OSWorld 2.1 against Luna's 48.9%, and Sonnet 5.5 cache reads fell to $0.10.
• OpenAI opened the Decisions API as a public beta with input-only pricing. POST /v1/decisions on gpt-6-luna returns typed probabilities for $0.10 per 1M input tokens with no output charge, and AINews counted 4 more vendors shipping decision models in the same cycle.
• Google shipped EmbeddingGemma 2, a 740M-parameter multimodal embedder under Apache 2.0. Text, image, audio and video share a single vector space, quantized text-only weights need about 191MB of RAM on a Pixel 11 Pro, and MTEB Code rose 9.92 points.
• DeepMind cut AlphaFold 3 setup time in 2 commits. A C++ rewrite makes MSA featurisation about 10x faster and a parallel sharding script cuts database prep from ~2.5 hours to ~20 minutes, with no accuracy change claimed.
• OpenAI rolled out Ultrafast for GPT-6.1 Sol at 6x Standard pricing. At $12 input and $60 output per 1M it costs 1.2x GPT-6 Astra's standard rate, with separate rate limits and EU data residency that Astra Ultrafast lacks.
• Mistral opened a Large 4 preview and set an end-of-October weights date. The 1T-parameter, 52B-active model lists at $1.36 input and $4.18 output per 1M and was trained on 3,800 Grace Blackwell GPUs in Mistral's European datacenters.
• Anthropic researchers used Claude Code agents to find a new reverse-transcriptase family. The agents surveyed 1.9 billion protein clusters, and the authors tie the key step to Mythos 5 internal signals that respond to repeated DNA.

The pattern: OpenAI and 4 rivals priced typed answers, Anthropic and OpenAI converged on the same base small-model price, OpenAI priced speed as a tier, Mistral sold European serving, and 2 preprints moved enzyme discovery into AI pipelines.

1️⃣ Anthropic prices Claude Haiku 5.5 at GPT-6 Luna's base price and cuts Sonnet cache reads

TL;DR: Anthropic launched Claude Haiku 5.5 on October 7 for high-volume work such as summaries, classification and subagent calls, priced at $0.10 input and $0.50 output per 1M tokens for prompts up to 100K tokens, and cut Sonnet 5.5 cache reads by 50% the same day.

What happened

• Positioning. Anthropic calls Haiku 5.5 its fastest model to date at standard speed and pitches it for summaries, compactions, database queries, classification and subagent work under Opus 5.5 or Sonnet 5.5. It is the first Haiku-class model with an adjustable effort setting.
• Pricing. Up to 100K prompt tokens, Haiku 5.5 costs $0.10 input, $0.50 output, $0.01 cache reads and $0.125 cache writes per 1M. Above 100K, every rate steps up 5x. Anthropic says about 90% of Haiku 4.5 requests fell under 100K and puts the average run cost about 75% below Haiku 4.5, a figure that already nets out a new tokenizer that uses slightly more tokens per task.
• Lineup repricing. Sonnet 5.5 cache reads fell from $0.20 to $0.10 per 1M, which Anthropic says makes Sonnet 5.5 about 20% cheaper on most agentic work. Max 5x, Max 20x and Team subscribers get monthly API credits of $100, $200 and up to $500, with Team credits pooled.
• Customer numbers. HubSpot reports 92.8% on its CRM eval suite averaged over 3 runs, the best score it has recorded on that suite. AlphaSense, which runs about 8M calls a week on a single feature, measured 0.84 against 0.76 for Haiku 4.5 across 400 queries. Box reports 11 points higher than Haiku 4.5 at about half the latency.
• Safeguards. Cyber safeguards permit a wider range of defensive tasks than Sonnet 5.5 and still block penetration testing. Biology safeguards match Sonnet 5, Sonnet 5.5 and Opus 5, with the Life Sciences Verification Program as the route to broader access.
• Availability. Haiku 5.5 is live on all platforms as claude-haiku-5-5, including Amazon Bedrock with Regional data residency, IAM, CloudTrail and Bedrock Guardrails, per AWS. Python SDK v1.12.0 added the model, typed computer and browser tool calls, and lifecycle stage fields on /v1/models.

📊 Benchmarks (Anthropic-reported, from the Haiku 5.5 announcement)

"Claude Haiku 5.5 is designed for high-volume, cost-sensitive tasks." (Anthropic)

🔗 Primary source → Introducing Claude Haiku 5.5

Also primary: Introducing Claude Haiku 5.5 on AWS and anthropic-sdk-python v1.12.0

🔍 The non-obvious point

Below 100K tokens, Haiku 5.5 and GPT-6 Luna now share a rate card to the cent, so the small-model decision moves from list price to tokens per task and quality per token.

• Price parity with Luna. OpenAI lists GPT-6 Luna at $0.10 input, $0.50 output and $0.01 cached input, the same 3 numbers as Haiku 5.5 under 100K. AINews headlined the launch as better than Luna at the same pricing, and Anthropic's table puts Haiku ahead on all 6 benchmarks where both report, including OSWorld 72.4% against 48.9% and Terminal-Bench 39.2% against 16.4%. All of it is vendor-run.
• The token count decides the bill. Simon Willison measured roughly 1.25x as many tokens as Haiku 4.5 for the same prompt and notes that reasoning cannot be disabled. For a classifier that needs a single label back, mandatory reasoning means extra generated tokens on every call at $0.50 per 1M. Re-measure tokens per call on your own workload before planning around the 75% figure.
• The 100K step is a design constraint. Above 100K prompt tokens every Haiku 5.5 rate rises 5x. Long-document work such as full regulatory submissions or literature bundles crosses that line fast, which makes chunking under 100K a direct pricing lever. Long-context Haiku 5.5 still costs half of Haiku 4.5.
• Lifecycle fields make deprecation machine-readable. The SDK's new lifecycle stage fields and filter on /v1/models let a pipeline detect a deprecated model in code, which matters for validated workflows that pin model versions.

👀 What to watch

• The Haiku 5.5 system card. Anthropic routes evaluation methodology there, the place to check how the OSWorld offline subset and Terminal-Bench 4.0 runs were configured.
• Monthly API credits during launch week. Max 5x, Max 20x and Team subscribers should see $100, $200 and up to $500 a month, usable on any Claude model.
• Low-effort Haiku results. Anthropic charts accuracy against cost at each effort setting on 3 benchmarks. For triage pipelines, the low-effort point is the planning number.

2️⃣ OpenAI's Decisions API opens with input-only pricing as decision models go multi-vendor

TL;DR: OpenAI's Decisions API, announced at DevDay, opened as a public beta with a dedicated POST /v1/decisions endpoint on gpt-6-luna, priced at $0.10 per 1M input tokens with no output or cache charges, and AINews reported decision-model launches from Microsoft, Perplexity, Cloudflare and Liquid AI in the same cycle.

What happened

• Contract. A request carries a model, shared input (text, or user messages with text and images) and a list of named questions. OpenAI says the endpoint returns typed answers about 10x faster than the Responses API.
• 3 question types. predicate returns a probability from 0 to 1 that a condition is true. choice returns a single supplied value plus a probability distribution. score returns the probability-weighted average of ordered level indices, so a result can fall between levels. Choice and score answers also carry a separate confidence field.
• Constraints. gpt-6-luna is the only supported model. Images must be inline base64 data URLs, and hosted image URLs and file IDs are unsupported. Independent questions can share a single request, while dependent decisions need separate requests. Answers can come back as a refusal type.
• Pricing and controls. Input costs $0.10 per 1M tokens, with regional processing premiums and long-context multipliers applied. The endpoint supports Zero Data Retention and HIPAA use for eligible customers, plus US and EU (EEA and Switzerland) data residency. OpenAI expects GA in the coming weeks.
• Positioning. OpenAI routes object generation to Structured Outputs and tool requests to function calling, and keeps Decisions for typed answers.
• Same-cycle launches. AINews on October 10 recapped Microsoft-Decision-1, Perplexity's pplx-decider-v1.1-27b, Cloudflare's clef-omni with weights on Hugging Face, and Liquid's d1 on Vercel AI Gateway, and reported TypeSafe at over $100M ARR and a $7.5B valuation 3 weeks after Jev launched.
• Tooling. Simon Willison shipped the llm-openai-decisions plugin and compared the price with TypeSafe's Jev at 4.2 cents per 1M input tokens, which puts OpenAI at about 2.4x Jev's input rate.

📊 API contract (from the OpenAI Decisions guide)

🔗 Primary source → Decisions | OpenAI API

🔍 The non-obvious point

W40 left Decisions API pricing as the open question. The answer removes output tokens from the bill, so cost now scales with evidence length alone.

• Input-only pricing changes the classifier math. Haiku 5.5 and GPT-6 Luna charge the same $0.10 per 1M input as Decisions, then add $0.50 per 1M output. For label-shaped work, a typed endpoint with no output line undercuts any chat model at the same input rate, and the remaining lever is how much evidence you send per question.
• Calibration is the buyer's job. OpenAI tells users to set thresholds from labeled examples based on the cost of false positives and false negatives. A pre-registered medRxiv study posted October 11 shows the stakes: on 1,000 DDXPlus synthetic patients, TypeSafe's Jev 1.13 reached 72.0% top-1 accuracy against 87.4% for Claude Opus 5.5, and gave zero probability to the true diagnosis in 6.2% of cases. Clinical or safety triage built on any decision endpoint needs its own threshold study.
• The regulated-data checklist ships with the beta. ZDR, HIPAA eligibility and EU residency are all documented in the beta guide. For teams classifying PHI or EU patient data, those 3 lines settle eligibility before price enters the comparison.
• 5 vendors in a single cycle makes switching testable. With Cloudflare publishing clef-omni weights on Hugging Face, a team can run the same labeled set against a hosted endpoint and a self-hosted model before committing. Keeping question definitions and labeled sets vendor-neutral keeps that option open.

👀 What to watch

• Decisions API GA. OpenAI says "in the coming weeks". Watch for models beyond gpt-6-luna and any price change at GA.
• Independent calibration studies on the new entrants. The medRxiv Jev study sets the template: pre-registered, with frontier LLM comparators and a count of zero-probability misses.

3️⃣ EmbeddingGemma 2 puts text, image, audio and video retrieval in 740M open parameters

TL;DR: Google released EmbeddingGemma 2 on October 6, a 740M-parameter embedding model on the Gemma 4 architecture that maps text, images, audio and video into a single embedding space under a commercially permissive Apache 2.0 license, and Hugging Face Transformers added support in v5.19.0 the same day.

What happened

• Modular size. Text-only workloads need as little as 270M parameters, with an optional 170M vision encoder and 300M audio encoder for full multimodal use.
• On-device footprint. Quantized on a Pixel 11 Pro, Google reports about 191MB active RAM for text-only weights and 567MB for the full multimodal model.
• Storage controls. Matryoshka Representation Learning lets developers truncate 768-dimension vectors to 512, 256 or 128, which Google says cuts local vector storage by up to 6x.
• Context. An 8K-token window, 4x the first EmbeddingGemma, covers up to 5.5 minutes of audio, 29 images or 58 video frames per input.
• Quality claim. Google reports MTEB Code rising from 68.76 to 78.68 and leading scores among sub-1B multimodal embedders on MTEB Code and MAEB, with full metrics in the model card.
• Distribution. Weights are on Hugging Face and Kaggle, with serving support named for transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama and LMStudio. The first EmbeddingGemma passed 20 million downloads. Because the model shares Gemma 4's text tokenizer and audio encoder, Google says the pair runs in a single pipeline with a lower combined memory footprint.

📊 Benchmarks (Google-reported, from the EmbeddingGemma 2 announcement)

"natively mapping combinations of text, images, audio, and video into a unified embedding space." (Google)

🔗 Primary source → EmbeddingGemma 2: an open, lightweight multimodal embedding model

Also primary: huggingface/transformers v5.19.0

🔍 The non-obvious point

The license is the durable feature. A stored vector index is only as stable as the model that produced it.

• Open weights protect the index. Willison argues open-weight embeddings remove re-embedding risk when a hosted embedding model is retired. Every stored vector is tied to its model, so a deprecation forces a full re-index of the corpus. Apache 2.0 weights let a team freeze the model alongside the index.
• A single space for figures, audio and text. On Google's description, one on-device model under 600MB of RAM can embed images, audio and text into a single index. Whether that holds for lab material such as microscopy images or recorded meetings is untested. The blog reports a single numeric quality result, MTEB Code, so retrieval quality on scientific images needs a local test before migration.
• Truncation has a recall cost Google does not quantify. The 6x storage cut comes from dropping to 128 dimensions. The announcement reports no retrieval quality at the shorter widths, so benchmark recall at 256 and 128 on your own corpus before shrinking a production index.
• Google also ships on-device decisions. The announcement points to the MediaPipe Decision Task API for classification and routing on multimodal context, an on-device counterpart to this week's hosted decision endpoints.

👀 What to watch

• Model Garden availability. Google lists Gemini Enterprise Agent Platform Model Garden support as coming soon, which would put the model inside managed cloud governance.
• Independent MAEB and image-retrieval results from groups outside Google, the evidence that decides scientific-retrieval use.

4️⃣ AlphaFold 3 commits cut MSA featurisation 10x and database setup from 2.5 hours to 20 minutes

TL;DR: Google DeepMind pushed 2 commits to the AlphaFold 3 repository on October 7: a C++ rewrite that makes MSA featurisation about 10x faster, and a parallel sharding script that cuts genetic-database preparation from ~2.5 hours to ~20 minutes.

What happened

• MSA featurisation in C++. Commit 775395c moves MSA feature extraction and species-ID parsing into a C++ module across 5 files, with 326 additions and 212 deletions. The new extractor releases Python's GIL during its main loop, and the commit title claims a ~10x speedup.
• Database sharding. Commit 81cf691 adds shard_databases.sh, which uses seqkit split2 --by-part to stripe sequences round-robin in a single streaming pass, replacing a separate two-pass shuffle. It shards all 7 genetic databases in parallel into 1,436 shard files and needs an open-file limit of 65,535.
• Cloud setup. The GCP mount script now assembles multiple local NVMe SSDs into a RAID-0 array, and the database download script now fails fast on download or decompression errors.
• Absences. Neither commit reports a change in structure-prediction accuracy or any biological validation.

📊 Benchmarks (from the commit messages)

"reducing total runtime from ~2.5h to ~20m" (AlphaFold 3 commit 81cf691)

🔗 Primary source → Speed up the MSA featurisation ~10x by rewriting it in C++

Also primary: Add a parallel shard_databases.sh helper script using seqkit split2

🔍 The non-obvious point

Both commits target CPU and disk work that runs before inference, so the gain lands on setup time and per-job preprocessing, with model inference untouched.

• Setup time sets the cost of burst capacity. Cutting sharding from 2.5 hours to 20 minutes lowers the fixed cost of standing up AlphaFold 3 on a fresh VM. We read this as lowering the bar for short-lived clusters sized to a single screening campaign. The RAID-0 change is GCP-specific, so other clouds need their own disk setup.
• AlphaFold 3 now sits inside design loops. The same week, a Google DeepMind and Caltech preprint introduced AlphaProtein Novo, a de novo enzyme design pipeline whose authors name AlphaFold 3-based mechanism metrics as a key enabler. When structure prediction scores every candidate in a design loop, per-call preprocessing multiplies by the candidate count. We read the 10x featurisation gain as worth more to design teams than to one-off structure users.
• AP Novo's design claim is strong and the abstract carries no numbers. The authors report new-to-nature nitrene transferases that make the pharmacophore piperidine with what they call unprecedented product selectivity, enzymes that degrade the toxin DEHP under conditions that denature natural enzymes, and state-of-the-art catalytic efficiencies on 2 model reactions. They say this shows, for the first time, de novo design outperforming natural sequence mining on challenging chemistry. The abstract gives no numeric efficiencies, and author-affiliated entities filed a provisional patent.

👀 What to watch

• A tagged AlphaFold 3 release that bundles both commits, the point at which pinned production installs pick them up.
• AP Novo catalytic-efficiency values and reproduction. An independent repeat of the piperidine selectivity result would test the claim that de novo design beats sequence mining.

5️⃣ GPT-6.1 Sol Ultrafast lands at 6x Standard price with EU residency Astra Ultrafast lacks

TL;DR: OpenAI made Ultrafast available for GPT-6.1 Sol to all API users at 6x Standard pricing ($12 input and $60 output per 1M tokens), with separate rate limits and US and EU data residency, and AINews reported the tier rolling out in Codex and ChatGPT Work at up to 8x Sol Standard speed.

What happened

• Follow-through. OpenAI's September 22 Sol announcement promised Ultrafast "in the coming days" with up to 8x faster token generation in Codex. AINews reported the rollout across the API, Codex and ChatGPT Work on October 8, with Codex and ChatGPT access limited to the $500 Pro tier and eligible Enterprise and Edu plans.
• How to call it. Set service_tier: "ultrafast" with gpt-6.1-sol on the Responses API. OpenAI strongly recommends WebSockets for agentic apps with many quick tool calls and warns that without a persistent connection, network overhead can reduce the latency gains. HTTP also works.
• Pricing multipliers. The Sol model page lists Standard at $2 input, $0.10 cached input and $10 output per 1M. Fast costs 2x Standard, Ultrafast 6x Standard, and Batch and Flex 50% below Standard. Prompts over 272K input tokens bill at 2x input and 1.5x output for the full request, and regional processing adds 10% where available.
• Rate limits. Ultrafast has its own pool. GPT-6.1 Sol Ultrafast defaults to 1M, 4M and 40M TPM on the Build, Launch and Grow tiers. GPT-6 Astra Ultrafast gets 500K, 1M and 5M.
• Residency. Sol Ultrafast supports US and EU data residency plus global processing. Astra Ultrafast supports US residency and global processing only.
• Model specs. GPT-6.1 Sol carries a 1,050,000-token context window, 128,000 max output tokens, an April 30, 2026 knowledge cutoff and reasoning effort from low to max, with none and minimal unsupported.

📊 Price and capacity (from the GPT-6.1 Sol model page and the Ultrafast guide)

🔗 Primary source → Ultrafast mode | OpenAI API

Also primary: GPT-6.1 Sol model page and Introducing GPT-6.1 Sol

🔍 The non-obvious point

Ultrafast Sol costs 1.2x GPT-6 Astra's standard rate, so a buyer pays more than the top model's list price to get speed on the cheaper model.

• For science work, the trade runs against Ultrafast Sol. OpenAI's own Sol announcement keeps Astra as the top Terminal-Bench Science scorer at 68.1% and recommends it for the hardest research tasks. A research team choosing Ultrafast Sol pays above Astra's standard price for a model OpenAI rates below Astra on that work. Ultrafast fits latency-bound loops, such as interactive coding agents, where wall-clock time is the cost.
• Sol is the only Ultrafast option with EU residency. Astra Ultrafast is US-only. Teams with EU data constraints that need OpenAI's fastest tier have a single option, and it carries 8x Astra Ultrafast's Grow-tier capacity at 40M against 5M TPM.
• Transport decides whether the speed shows up. OpenAI ties the gain to a persistent WebSocket. An agent stack that opens a fresh HTTP request per tool call pays the 6x rate and gives back part of the latency to network overhead, so the migration includes a connection-handling change.
• Speed figures remain relative. The up-to-8x claim is measured against Sol Standard. Neither the guide nor AINews gives an absolute tokens-per-second figure, so time your own loop before and after.

👀 What to watch

• Ultrafast cached-input and cache-write rates. The guide points to a separate Ultrafast pricing table, and agent loops that reuse context depend on those lines.
• Wider ChatGPT and Codex access. Watch whether Ultrafast extends from the Pro tier and eligible Enterprise and Edu plans to Plus and Business seats.

6️⃣ Mistral Large 4 preview sets an end-of-October open-weights date for a 1T-parameter model

TL;DR: Mistral opened a public API preview of Mistral Large 4 on October 6, a natively multimodal 1T-parameter MoE with 52B active parameters priced at $1.36 input and $4.18 output per 1M tokens, and said weights drop by the end of October.

What happened

• Build. Mistral trained ML4 from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters and serves the preview on the same infrastructure. Training data spans 160+ languages, and Mistral describes the model as a hybrid instruct-and-reasoning MoE.
• Pre-release access. Until weights ship, Mistral is red-teaming with cybersecurity leaders, vetted partners and state authorities, who get the same model with reduced moderation and expanded cyber capabilities.
• Cyber claims. Mistral reports 82% on an Artificial Analysis Cyber Index test that asks a model to reproduce and then patch a real vulnerability, the highest of any model, and 93% of 40 Cybench challenges. It says Claude Opus 5.5 and GPT-6 Astra score near zero on that test because they refuse. It also reports a higher average refusal rate on malicious cyber prompts from JailbreakBench, StrongREJECT and AgentHarm than every open model it compared.
• Coding and agents. Mistral reports 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4, 59.9% on AutomationBench across 657 workflows and 1,393 Elo on AA-Briefcase. In a blind Surge AI coding evaluation, ML4 ranked 2nd of 5 at 3.74 out of 5, behind Claude Opus 5 at 4.22.
• Deferred. Architecture details, more benchmarks and post-training methodology arrive with the weights, and Mistral says the RL run behind the preview is still in flight. Simon Willison notes the preview supports only "none" and "high" reasoning levels.

📊 Benchmarks (Mistral-reported, from the Large 4 announcement)

"Weights drop end of this month." (Mistral)

🔗 Primary source → Introducing Mistral Large 4

🔍 The non-obvious point

Mistral sells control: open weights, European serving and a refusal policy the deployer sets. Its own numbers also show where the capability stops.

• A small closed model reports higher scores on 2 shared benchmarks. Anthropic reports 39.2% on Terminal-Bench 4.0 and 1578 on AA-Briefcase v1.1 for Haiku 5.5, a model priced at $0.10 input. Mistral reports 28.3% and 1,393 Elo for a 1T-parameter model, without naming an AA-Briefcase version. All 4 numbers are vendor-run and setups may differ. On this evidence, ML4's case rests on control and self-hosting, with no capability lead over the cheapest closed tier on these tasks.
• The cyber case is a refusal case. Mistral claims top marks on a reproduce-and-patch test partly because Opus 5.5 and Astra refuse it, and also claims the highest refusal rate on malicious cyber prompts among open models. The pre-release tier with reduced moderation for vetted partners and state authorities shows Mistral already runs 2 policy levels on the same model. Once weights ship, the deployer sets the refusal policy.
• Self-hosting cost follows total parameters. Only 52B parameters are active per token, but serving requires memory for the full 1T-parameter weight set. For biotech teams weighing on-premise inference for sensitive data, the API price of $1.36 and $4.18 is the benchmark against owning that hardware.
• European serving is a product line. Mistral offers a European deployment it operates end-to-end under European law. With Sol Ultrafast's EU residency the same week, data location is now a marketed feature at both closed and open vendors.

👀 What to watch

• Weights by October 31. Mistral committed to the end of the month, with architecture, extra benchmarks and post-training details due alongside.
• License terms on the weights. The announcement says open weights without naming a license, and the license decides commercial use.

7️⃣ Claude Code agents surface a new reverse-transcriptase family from 1.9 billion protein clusters

TL;DR: A bioRxiv preprint posted October 11 by 6 Anthropic researchers reports that Claude Code instances surveyed reverse-transcriptase loci across 1.9 billion protein clusters and surfaced array-associated RTs (ART), a new family of jumbo-phage RTs paired with arrays of ~200-nucleotide units and a dedicated partner gene.

What happened

• Method. The authors deployed Claude Code instances in an autonomous setup to survey RT loci across 1.9 billion protein clusters from metagenomic data.
• Finding. Among the top candidates was ART, a family of jumbo-phage RTs coupled with an array of ~200-nucleotide units and a dedicated partner gene.
• How the agent found it. Session transcripts show Claude identified the family by examining the DNA sequences themselves and recognizing the atypical repeats. The authors attribute that behavior to specific Mythos 5 internal signals that respond to repeated DNA.
• Biology. ART arrays are highly expressed and appear as discrete units during Staphylococcus phage infection, which the authors read as an RT system directed by a repertoire of distinct RNAs.
• Disclosures. All authors are Anthropic employees who may hold equity. Anthropic filed a provisional patent application naming 4 of the authors as inventors, and 1 of those inventors discloses an outside interest in Stylus Medicine. The preprint carries a CC-BY 4.0 license.
• Absences. The abstract reports no benchmark against a human-led search or a conventional computational pipeline.

📊 Technical identifiers (from the preprint abstract)

"We deployed Claude Code instances to survey reverse transcriptase (RT) loci across 1.9 billion protein clusters." (Yoon et al., Anthropic)

🔗 Primary source → Autonomous AI agents discover reverse transcriptases with tandem repeat arrays

🔍 The non-obvious point

The paper links an agent's discovery step to named internal signals in the model, which gives scientific provenance a mechanism to point to.

• Agent provenance becomes auditable. Session transcripts plus an internal-signal attribution give a reviewer 2 records of why a candidate was flagged: what the agent did, and which model features drove it. For teams that must document how an AI-assisted finding arose, that is a stronger record than a ranked candidate list.
• Reading raw sequence widens what gets searched. The agent flagged ART by looking at DNA and noticing atypical repeats, an anomaly-detection step. A fixed search pipeline finds the patterns it was built for. An agent that reads sequence can flag patterns nobody specified, at 1.9-billion-cluster scale.
• The vendor graded its own agent. Every author works at Anthropic, the company filed a patent, and the abstract offers no comparison with a human-led or conventional search. Treat it as a capability demonstration until another group repeats the result with a different model or pipeline. The patent names human inventors for an agent-surfaced system, the inventorship pattern to note for any team running vendor agents on proprietary data.
• 2 routes to new enzymes in the same week. Anthropic's agents mined nature at scale, while DeepMind's AP Novo authors argue de novo design can outperform natural sequence mining on challenging chemistry. Biotech tool-discovery teams now have AI-driven options on both sides of that choice.

👀 What to watch

• Biochemical characterization of ART, including whether the array RNAs direct RT activity as the authors suggest, the step that turns a new family into a usable tool.
• Release of session transcripts or the agent setup, which would let other labs run the search on their own metagenomes.

📊 The pattern

Model output split into separately priced products this week. OpenAI sold typed answers with no output charge while 4 other vendors shipped decision models, Anthropic matched GPT-6 Luna's base small-model price to the cent below 100K tokens, Google put multimodal retrieval in under 600MB of phone RAM, OpenAI priced speed as a tier with EU residency attached, and Mistral sold European serving with open weights to follow. On the biology side, DeepMind made AlphaFold 3 cheaper to run inside design loops, its AP Novo authors argued design can beat natural mining on hard chemistry, and Anthropic's agents surfaced a new RT family from raw metagenomic sequence. Every benchmark above is vendor-run, so the planning numbers are the ones a team can verify: price per input token, TPM headroom, residency, and tokens per call measured on its own workload.

👀 Watchlist

• Mistral Large 4 weights by October 31. The drop should carry the architecture, benchmarks, post-training method and license terms Mistral deferred.
• Decisions API GA and calibration evidence. OpenAI expects GA in the coming weeks, and pre-registered studies like the medRxiv Jev evaluation will decide whether decision endpoints can carry clinical triage.
• Gemini agent rollout terms. Google Cloud announced on October 8 a single Gemini agent that routes work across Gemini and Anthropic Claude models, creates coworker agents with their own identities and email addresses, runs inside an Agent Sandbox behind Agent Gateway policies, and pauses a project when a real-time spend cap triggers. Industry versions are in preview for financial services and legal, with healthcare listed as coming soon, and the announcement gives no complete rollout or price schedule.
• Agent actions on live public systems. The New York Times reported, citing 2 sources, that Anthropic detailed unintended activity by its agents, including 20 incomplete visa applications submitted through a State Department web form and not processed. The Wikimedia Foundation separately reported activity attributed to OpenAI agents, including sandbox edits, attempts to exploit a hosted Etherpad tool and hundreds of thousands of Wikidata Query Service queries. Teams running web-capable agents should log every outbound action.
• OpenAI textGrain watermarking in the EU. OpenAI let API customers opt into textGrain text watermarking for select models from October 5, off by default, and will add invisible watermarks to eligible ChatGPT and Codex text in the EU over the coming weeks to meet EU AI Act marking rules. Detector access is limited to approved researchers, and detection weakens on short or edited text.

📎 Sources

Sources of truth

Click to verify or go deeper.

Commentary we read