LLM News Today: Flagships, Open Weights and Falling Prices

The short version of LLM news today is that the frontier moved three times in two months. Anthropic rebuilt its whole Claude ladder between September 1 and October 7. OpenAI rolled out the GPT-6 family and then put those models on all three big clouds. Google took Gemini 4 Argon fully public on October 13. In the same weeks, downloadable models from Meta and Mistral reached scores that closed flagships held a year ago, and the two highest ranked open models on the independent index now come from China.

An LLM, short for large language model, is a very large neural network trained on text until it can predict and produce language. That one idea now sits inside chatbots, coding tools, search boxes and office software, which is why model releases have become weekly news. One release tracker counted 46 model releases from 27 makers in September 2026 alone. Set that pace beside the AI news of December 2025, when a single flagship launch could lead a whole month, and the speed of this field becomes clear.

The flagship race: three ladders rebuilt in eight weeks

Anthropic moved first and fastest. Claude Fable 5.1 arrived on September 1 as the new top model, with a restricted twin, Mythos 5.1, held for approved customers only. Opus 5.5 followed on September 22, priced at $4 per million input tokens and $20 per million output tokens, about 20 percent below Opus 5 per token, with the cost of reading cached input cut by 60 percent. Anthropic says Opus 5.5 matches Fable 5.1 on most work. Sonnet 5.5 landed on September 28 at $2 and $10, with output that Anthropic says runs more than 30 percent faster than Sonnet 5. Haiku 5.5 closed the ladder on October 7, adding a one million token context window and a running cost Anthropic puts at about 75 percent below the Haiku it replaces, as 9to5Mac reported on launch day. Every performance claim in that list comes from Anthropic’s own testing, a pattern worth noticing wherever it appears.

OpenAI spread the GPT-6 family across the autumn. GPT-6 Astra, the top tier aimed at coding, computer use and scientific research, became generally available on Amazon Bedrock on September 8. A 6.1 refresh of Sol, pitched as near-Astra intelligence for less money, followed in late September, with Luna as the fast and cheap tier for everyday work at scale. The mid-priced Terra tier still runs the older 5.6 generation, which makes the lineup a useful reminder that a new family name does not mean every tier moved at once.

Google announced Gemini 4 Argon on September 30 and gave first access to cyber defenders in its Fairwind program, a staged release pattern that has become normal for powerful models. Argon reached general availability in the Gemini API on October 13, with a context window just over one million tokens and launch prices of $2 per million input tokens and $10 per million output tokens. Those prices carry a deadline, covered in the money section below. Meta split its bet in two: Muse Wonder, its strongest model, launched as a closed, API-only flagship on October 12, and its open sibling followed three days later. These flagships are also the engines inside the agent tools now spreading through offices, a shift we track in our multi-agent AI news coverage.

Step back from the names and the shared spec sheet of LLM news today comes into focus. A context window of about one million tokens, once a headline feature, is now standard equipment: Argon, Opus 5.5 and Spark Open all list it, and AWS lists Astra with up to one million input tokens. Adjustable thinking time is the second shared feature, letting buyers trade speed and cost against depth of reasoning inside the same model. The third is the staged release itself: defenders and approved partners first, the public a few weeks later.

Open weights: the gap is now a few points

The biggest change in LLM news today sits in the open-weight race, the contest between models anyone can download and run. Meta published Muse Spark Open on October 15: 816 billion total parameters, with 42 billion active on any given token, a one million token context window, and input that covers text, images, video and audio. Meta calls it the most capable US open-weight model to date. On the Artificial Analysis Intelligence Index, an independent scoreboard that reruns models under one set of rules, Spark Open scores 39. That leads every US open model and still trails two Chinese releases: GLM 5.3 from Z.ai at 45 and Kimi K3 from Moonshot AI at 44.

France answers with Mistral Large 4, a model of about one trillion parameters that entered public preview through Mistral’s API on October 6. It scores 38 on the same independent index, up from a score of 9 for Mistral Large 3, and the company has promised the downloadable weights by the end of October. Mistral’s own announcement frames the release as closing the gap with the best Chinese open models. China, meanwhile, keeps its lead the plain way: DeepSeek folded its planned separate reasoning model into a single V4 family and shipped V4.1-Flash in September, tuned for coding agents. The pattern across all four makers is identical: publish the weights, let anyone download them, and compete on what independent testers measure. We follow the download pages, licenses and leaderboards in our Hugging Face news coverage, because that is where open releases land first.

Licenses decide what open really means after the download. Spark Open ships under Meta’s new Muse Model License, which allows commercial use but adds terms the older Llama license never had, including a revenue clause for anyone selling hosted access to the model. Glimmer, by contrast, uses Apache 2.0, the permissive license with the fewest strings attached. Mistral has not published Large 4’s final terms yet. Two models can both be called open while giving builders very different rights, which is why the license deserves the same reading time as the benchmark.

Local LLMs: open weights are not the same as runs on your desk

Local LLM news today carries a catch that download pages rarely shout about. Spark Open is free to download, and the weights are genuinely public. The smallest usable compressed version still needs about 180 gigabytes of memory, and the setup the tooling recommends starts near 250 gigabytes. No normal laptop and no gaming PC comes close. On Ollama, the popular tool for running models at home, Spark Open exists only as a cloud tag, which means it runs on Ollama’s servers rather than on your machine. Open weights and runs locally are two different promises, and this season’s biggest release only makes the first one. The tooling around the smaller models is the mature part of this story: the GGUF compressed files and llama.cpp support that local runners depend on now arrive on launch day instead of weeks later.

The honest local story of the season is smaller. Meta’s Muse Glimmer 30B, released on August 10 under the permissive Apache 2.0 license, compresses to about 17 gigabytes and runs on a secondhand RTX 3090 graphics card, the kind of hardware a hobbyist can actually buy. Mistral went smaller still on October 15 with Timestral, a family of forecasting models that needs no graphics card at all: the largest holds 209 million parameters and produces a forecast on a plain server CPU. Mistral says it beats the previous best specialist forecasters on three public benchmarks, and it submitted those results to the public leaderboards instead of asking readers to simply believe them. Small, narrow and checkable is a real direction in LLM research news, not a footnote to the flagship race.

Benchmark news: read who ran the test

The safest habit in LLM news today is to ask one question of every chart: who ran this test? Launch posts almost always report the maker’s own measurements, taken with the maker’s own prompts, tools and time limits. Scores also rise when a model is allowed to think longer before answering, and that longer thinking burns more tokens, so a headline score can hide a much higher cost per finished task. Two labs can test the same rival model under different rules and publish different numbers, and both numbers can be honest.

Independent testers exist to flatten those differences. Artificial Analysis reruns every model under the same rules and publishes one index, which is why its open-weight table, shown above, puts Chinese models on top while US launch posts lead with US models. Anthropic’s speed claim for Sonnet 5.5 and Meta’s most capable claim for Spark Open may both hold up, and both arrived as the makers’ own words. Treat every launch chart as an opening argument, then check the independent scoreboard a week later before you switch tools or sign a contract. LLM evaluation news moves slower than launch news, and it is the part that decides what the launches were actually worth.

Prices and access: the quiet stories move real money

Three money stories in LLM news today matter more than most launches. First, Google’s launch price for Gemini 4 Argon ends on November 30: the $2 and $10 rates per million tokens double to $4 and $20 on December 1, so any budget built on October prices quietly breaks in December. Second, Anthropic spent the autumn cutting the cost of cached input, the repeated context that apps resend with every request: cache reads fell 60 percent on Opus 5.5 at launch, then halved again on Sonnet 5.5 in October. Cache pricing never makes headlines and decides real bills. Third, access widened: when OpenAI models entered Google Cloud’s model garden on October 21, a company could run GPT, Claude and Gemini models under one bill for the first time, as Google’s own announcement explains.

One access story deserves a warning label. On October 8, Google deprecated its Gemini 3.7 Flash and 3.5 Flash models, and its release notes say that calls addressed to those names now route silently to newer Flash models. A team that pinned a model name in its code, expecting identical behavior, can wake up running a different model with different answers and different costs. The fix is boring and essential: pin exact model versions wherever the provider allows it, and read deprecation notes with the same care you give to bills.

What to watch next

LLM updates today arrive weekly, and the next few weeks carry four dated tests of everything above. Mistral’s Large 4 weights are due by the end of October, and the license attached to them will say more about the open race than the parameter count does. The Argon price rise lands on December 1 and will show whether Google’s flagship keeps its new users at full price. The independent index gets a new open-weight challenger judged against GLM 5.3 before it is judged against anything else. And OpenAI still owes its mid Terra tier a 6.1 refresh, the one visibly unfinished piece of the GPT-6 rollout.

One more pattern in LLM news today deserves attention: the specialists are multiplying. Timestral does one job, forecasting, at a size that fits on a CPU. Coding-tuned releases like DeepSeek’s V4.1-Flash do one job inside agent tools. The general flagship still leads the news, but a growing share of real work is quietly moving to small models built for a single task, priced like utilities instead of like supercomputers.

Conclusion

LLM news today describes a field that no longer moves in yearly generations. The three big closed labs rebuilt their flagship ladders inside eight weeks. Open-weight models, led for now by two Chinese releases, sit a few independent points behind the best downloadable US and European entries, and the honest local models are the small ones. Prices fell at almost every launch, and the two exceptions, a scheduled doubling and a silent rerouting, both hide in fine print. Read the dates, ask who ran the benchmark, and check the price page before the launch post. We will keep the dates and sources on this page current as the next releases land.