The LLM Frontier in September
Price cuts in the cloud, 30B models locally, Jev on the gateways

September was less about who has the single best model and more about what that intelligence costs, and where it can run. In the cloud, mid-tier models got 50% cheaper. Locally, 30B-class models that fit in about 20 GB are becoming the default for agents. And Jev, a model that returns decisions instead of text, can now be called from mainstream developer platforms.

1. Cloud LLMs: the top tier holds, the middle gets much cheaper

OpenAI GPT-6 Sol / Luna (Sep 22)Verified

OpenAI released its flagship GPT-6 Astra in early September. On September 22 it added the mid-tier GPT-6 Sol and the lightweight GPT-6 Luna, both trained with methods similar to Astra's. API prices are 50% lower than GPT-5.6's promotional pricing. Cached input reads get a 90% discount, and OpenAI added a dashboard and diagnostics tool for checking how well caching is working. Source: OpenAI

Model	Input (per 1M tokens)	Output (per 1M tokens)	vs. previous generation
GPT-6 Sol	$2	$10	Half of GPT-5.6 Sol ($4 / $20)
GPT-6 Luna	$0.10	$0.50	Down from GPT-5.6 Luna ($0.20 / $1.20)

On AutomationBench, a benchmark of business workflows, Sol at its highest effort setting (xhigh) scored 33.2% at $0.27 per task. OpenAI says this beats Claude Opus 5 at max effort (26.9%) for about 9% of the cost.Vendor claim These comparisons are OpenAI's own, and competitor figures come from public reports. Note that Anthropic's Opus 5.5, released the same day, was reported at 40.0% on AutomationBench in Zapier's own evaluation, so the comparison was out of date within hours.

Google Gemini 3.8 Flash / 3.8 Flash Cyber (Sep 2)Verified

This is Google's third Flash release in six weeks. Gemini 3.8 Flash improves long-horizon coding and agentic work while keeping 3.7 Flash's introductory price ($0.75 input, $3.75 output per 1M tokens). 3.8 Flash Cyber, specialized for finding and patching vulnerabilities, is available only to trusted defenders through the new Fairwind Program. Google says its Chrome Security team got 2.6 times as many correct patches as with the best, much larger commercial models.Vendor claim Source: Google

One point stands out: Google says 3.8 Flash "works harder" and may use more tokens, especially at higher effort levels. So even at the same price per token, the cost per task can go up. For efficiency-first workloads, Google still supports 3.7 Flash.

Anthropic Claude Fable 5.1 / Mythos 5.1 (Sep 1) and Opus 5.5 (Sep 22)Verified

On September 1, Anthropic released Fable 5.1 and Mythos 5.1. They are the same model with different levels of safeguards. The general-release Fable 5.1 can find software vulnerabilities but cannot develop exploits for them, and its advanced biology research capabilities are restricted. Fable 5.1 is generally available, while Mythos 5.1 is offered only through trusted-access programs for cybersecurity and life-sciences work. Base prices stay at $10 input and $50 output per 1M tokens, but cache reads drop from $1 to $0.25 (confirmed on the pricing page). As a result, Anthropic estimates Fable 5.1 will cost about 25% less than Fable 5 on typical token-billed workloads, and up to about 45% less on highly agentic work. Source: Anthropic

Anthropic also reported results in scientific research.Vendor claim

• Protein design: Two external organizations tested binders designed by Mythos 5.1 in the lab. Nearly 50% of designs bound their targets across 12 targets, compared with typical hit rates of 10–15%.
• Venus elevation map: Fable 5.1 produced a new high-resolution elevation map covering a third of Venus, released under a Creative Commons license.
• Faster computational biology: Mythos 5.1 sped up seven open-source genomics and protein models by up to 2.5 times with identical outputs.

On September 22, Anthropic released Opus 5.5, the first model in the Claude 5.5 family. It costs $4 input, $20 output and $0.20 for cache reads per 1M tokens. Anthropic says it matches Fable 5.1 on most work and costs about 40% less than Opus 5 on typical workloads. Source: Anthropic

2. Local LLMs: 30B models in about 20 GB become the standard for on-device agents

Qwen3.8-27B (released Aug 14)Verified

Alibaba's Qwen team released a 27B-parameter vision-capable model under Apache 2.0. Simon Willison calls 27B an excellent size for a reasonably specced laptop, but notes that by default the model tends to overthink. Hugging Face / Simon Willison

Meta Muse Glimmer (Aug 10)Verified

After moving to a closed model with Muse Spark, Meta returned to open weights with a 30B agentic model under Apache 2.0, trained by distillation from Muse Spark's outputs. Quantized to about 4 bits, the language model fits in under 20 GB. That leaves room within 24–32 GB of memory for the image encoder and a speculative-decoding drafter. Meta reports decode speedups from the drafter of 3.1x on an RTX 5090 and 1.8x on an M5 Max.Vendor claim Source: Meta

3. Jev update: on major developer platforms the day after launch

TypeSafe AI announced Jev on September 15. On September 16 it became available on Vercel AI Gateway, and it is also listed in Cloudflare Workers AI as typesafe/jev.Verified

Item	Details
How to call it	The experimental evaluate API in AI SDK 7 (7.0.105 or later), or Cloudflare's env.AI.run('typesafe/jev', …)
Question types	Choice (pick an option), Score (grade on an ordered rubric), Noul/Boolean (probability of true). Multiple questions are evaluated in parallel
Context window	32,000 tokens (per Cloudflare)
Data protection	Zero Data Retention and No Training can be set per request
Model version shown	jev-1.13.0 in Cloudflare's example responses

Vercel's announcement says TypeSafe measured Jev as up to 193.6x faster and 444.6x cheaper than LLMs on its own workflow evaluations.Vendor claim Vercel also tells users to calibrate probabilities and confidence against labeled examples from their own workflows. The intended pattern is to automate clear cases and send uncertain ones to human review. Source: Vercel / Cloudflare

4. Two papers on reasoning without words, and what they mean for AI for science

Today's reasoning models get better by writing out their thinking as text (chain of thought, or CoT). The two papers below point in a shared direction: keep the reasoning as continuous vectors inside the model instead of converting it into tokens (words).

Paper A: Coconut, a chain of continuous thought (Meta FAIR / UC San Diego)Verified

Coconut takes the model's last hidden state (its internal vector) and feeds it back in as the next input, without turning it into a word. This lets a single "thought" hold several candidate next steps at once. As a result, behavior resembling breadth-first search (BFS) emerged, even though the model was never explicitly trained to search.

Task	CoT (reasoning in text)	Coconut	Tokens generated (CoT → Coconut)
ProsQA (logic that requires planning and search)	77.5%	97.0%	49.4 → 14.2
ProntoQA (logical reasoning)	98.8%	99.8%	92.5 → 9.0
GSM8k (math word problems)	42.9%	34.1%	25.0 → 8.2

On tasks that require search, Coconut beat text-based reasoning while using far fewer tokens. On math, it still falls short of CoT, though it is well above answering with no reasoning at all (No-CoT, 16.5%). arXiv:2412.06769

Paper B: Recurrent Depth, a model that can think deeper after training (ELLIS / University of Maryland / Lawrence Livermore National Laboratory)Verified

This language model applies its core Transformer block repeatedly, so it can think more deeply at inference time simply by running more iterations. The authors trained a 3.5B-parameter model on 800 billion tokens using Oak Ridge National Laboratory's Frontier supercomputer (AMD MI250X GPUs). More iterations improved results on reasoning benchmarks, with gains continuing up to compute equivalent to a 50B-parameter model. It needs no special reasoning training data and works with short context windows.

The authors also report that several features come without extra training:

• Per-token adaptive compute: stop iterating early on easy tokens
• KV-cache sharing: reduce memory use
• Self-speculative decoding: use the model itself as a drafter to generate faster

When the researchers tracked the model's internal state, tokens involving numbers followed orbit-like paths in latent space. arXiv:2502.05171

What changes for AI for scienceAIfrontier view

Aspect	What it could enable	What to watch
Exploring hypotheses	Coconut's BFS-like parallel search suits retrosynthesis (working backward to find how to make a target compound), proof search and branching experiment design — problems where committing to one path too early leads to failure	It has only been tested on synthetic logic tasks. There is no evidence yet on real chemistry or physics data
Reasoning that is hard to put into words	The Recurrent Depth authors suggest latent space could handle reasoning that resists verbalization, such as spatial thinking and physical intuition. This might avoid information lost when molecular shapes or fields are forced into text	The authors present this as a hope, not a demonstrated result
Allocating compute	Iterations can be increased only for hard problems, so limited GPUs can be focused where they matter most. The design fits well with pre-screening and narrowing candidates before large simulations	More iterations mean more compute per problem, so cost-effectiveness must be measured task by task
Reproducibility and auditing	Not writing out the reasoning can cut token counts, and therefore cost, substantially	It becomes harder for people to read and check how a conclusion was reached, which matters most in science. CoT was an imperfect audit trail, and that trail is lost

By contrast, Anthropic's protein-design result (above) used an existing model that reasons in text, gave it design tools, and checked the results in external lab experiments. Today, what produces scientific results is this loop of using tools and verifying experimentally. The realistic role for latent reasoning is as a component that speeds up search inside that loop.

This is why we think latent reasoning in science will depend on pairing fast internal search with typed verification of the results. Candidates found in latent space would be narrowed by simulation, formal verification, or systems like Jev that return typed decisions with probabilities, and only low-confidence candidates would go to people and experiments. This follows the same direction as our discussion of calibration and decision models, which treats decisions as typed probabilities, and the emphasis on verifiable structure in Sign. This is an editorial hypothesis, not a proven method.

Sources

• OpenAI, "Introducing GPT-6 Sol and Luna" https://openai.com/index/introducing-gpt-6-sol-and-luna/
• Google, "Gemini 3.8 Flash and 3.8 Flash Cyber" https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/
• Anthropic, "Introducing Claude Fable 5.1 and Claude Mythos 5.1" https://www.anthropic.com/claude-fable-and-mythos-5-1
• Anthropic, "Introducing Claude Opus 5.5" https://www.anthropic.com/claude-opus-5-5
• Anthropic, "Model pricing" https://docs.anthropic.com/en/docs/about-claude/pricing
• Qwen3.8-27B (Hugging Face) https://huggingface.co/Qwen/Qwen3.8-27B
• Meta, "Introducing Muse Glimmer" https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model
• Vercel, "TypeSafe AI's Jev now available on AI Gateway" https://vercel.com/changelog/typesafe-ai-jev-now-available-on-ai-gateway
• Cloudflare, "Jev (typesafe)" https://developers.cloudflare.com/ai/models/typesafe/jev/
• Hao et al., "Training Large Language Models to Reason in a Continuous Latent Space," arXiv:2412.06769
• Geiping et al., "Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach," arXiv:2502.05171