UnsplashAI tooling in September 2026: two frontiers in three days
GPT-6 Astra and Claude Fable 5.1 landed 48 hours apart at the same price, August was the densest release month on record, and the open tier quietly caught up on context.
Both frontier labs shipped a new top model in the first three days of September, at the same headline price, 48 hours apart. Behind them, August delivered the densest run of open-weight releases the field has seen. Here is what actually changed for people building things.
The frontier moved twice in one week
Claude Fable 5.1 arrived on 1 September at $10 per million input tokens and $50 per million output. GPT-6 Astra followed on the 3rd at the same list price for short context, rising to $20 and $75 past the threshold. Both carry roughly a million tokens of context and cap output at 128K. On the face of it this is a tie.
The split is in what each is built to do. Astra is a generation change aimed at operating software: OpenAI positions it as state of the art on computer use, browsing and long end-to-end work, the kind of task that spans a spreadsheet, a browser tab and a terminal. Fable 5.1 is a point release aimed at thinking: Anthropic's own numbers put it at 52.6% against Opus 5's 29.0% on scientific research, and 73.4% against 70.0% on coding.
If you only change one thing
Check what your code defaults to, then check whether it should default lower rather than higher. Anthropic still recommends Opus 5 at $5 and $25 for most workloads, half the flagship price. And Fable 5.1 now bills cache reads at 2.5% of its input price rather than the usual 10%, so an agent that re-reads the same large context every step got materially cheaper without anyone changing a line.
One release did not go to everyone. Claude Mythos 5.1 is the same underlying model as Fable with additional safeguards for cybersecurity and life sciences work, and it is restricted to vetted organisations. Astra has a version of the same problem handled differently: it is the first model OpenAI has rated at a Critical cybersecurity capability level, and the version customers get refuses advanced cybersecurity tasks. Capability is starting to ship with an access-control story attached, which is new.
August was the densest month on the calendar
The open tier did most of the work. Qwen3.8 landed as a family rather than a model, the 2.4T-A95B flagship on the 8th, a 27B on the 13th, a Flash-Next variant on the 24th, all at 256K context. GLM-5.3 and GLM-5.3-Flash arrived from Z.ai on the 25th carrying a full million tokens. DeepSeek added a vision variant to the V4 Flash line at the end of the month.
The number worth noticing is context. A year ago the open models were the 128K tier and long context was the reason to pay for a closed one. GLM-5.3, DeepSeek V4 and Kimi K3 all sit at a million tokens now. That argument is over.
Meta shipped Muse Glimmer on 9 August, a 30B model under Apache 2.0, distilled from the closed Muse Spark and sized to run inside a 24GB VRAM envelope on a consumer GPU. It is worth being precise about what this means for Llama: Meta's last Llama weights are from April 2025, and its new open work now ships under a different name. If your local setup is still built around Llama out of habit, that habit is now about seventeen months old.
The interesting releases this month were not the biggest models. They were the ones small enough to run on the machine already on your desk.
Two standards landed, and they matter more than the models
MCP shipped its 2026-07-28 specification, the largest change since remote servers arrived. The protocol core is stateless now, sessions and the initialize handshake are gone, and a remote MCP server is an ordinary HTTP workload. Roots, Sampling and Logging are deprecated on a twelve-month clock, as is the legacy HTTP+SSE transport. If you maintain a server, this is the item on this page with a deadline attached, see what the new MCP spec changes.
Agent Plugins 1.0.0 was published on 6 August: one directory format that packages Agent Skills and MCP servers together, so an extension written once loads in more than one client. The steering committee is five people from Amazon, Cursor, Microsoft, OpenAI and Vercel. We wrote up what the format actually is, including the part most coverage skipped, which is that its security model is a filesystem containment rule rather than a permissions system.
Also worth noticing
- Gemini 3.8 Flash continues Google's habit of shipping a new Flash roughly every three weeks, at $0.75 and $3.75 per million tokens on introductory pricing that runs to 31 December 2026
- Anthropic previewed a Model Hardware Standard on 27 August, a common driver interface so agents can operate microscopes, liquid handlers and robotic arms without a vendor-specific integration each time
- Windsurf is now Devin Desktop. Cognition renamed it in June and both windsurf.com and codeium.com redirect. The editor is now the free entry point to Devin's paid cloud agent, which is a different product strategy from the one people signed up for
What to do about it
- Audit the model constant in your codebase, and consider moving down a tier rather than up. The mid tiers absorbed most of this year's gains
- If you run an MCP server, read the 2026-07-28 spec now. The deprecations have a twelve-month clock and stateless is a real migration
- If you self-host, re-benchmark. The open tier reached a million tokens of context this month and your last comparison is probably out of date
- Ignore the rest until something you are building actually needs it
We publish this monthly. The tool catalog, skills directory, MCP directory and model comparison are kept current between roundups. Prices here are list rates as of 4 September 2026 and they move often, confirm on the provider's page before you commit.
Find these on the Radar
Every tool here lives on Kapyn Radar. Save the ones that fit into a Loadout and find them again.
