Launches on 2026-09-22

Xiaomi's open-weight flagship: 1.02T parameters with 42B active, a 1M-token context and text, image, video and audio input, at $0.435 in and $0.87 out per million tokens — the price of the mimo-v2.5-pro that Xiaomi retires on 21 October.

The smaller MiMo-V2.6: 309B parameters with 15B active, the same 1M context and omni-modal input, open under MIT, at $0.14 in and $0.28 out per million tokens — and the replacement for mimo-v2.5 before 21 October.

Upstage's compact language model for agentic work, with 35B parameters (3B active), a 512K-token context and Korean, English and Japanese support, priced at $0.10 in and $0.40 out per million tokens.

The working tier of the GPT-6 family at half its predecessor’s price: $2 in and $10 out per million tokens, $0.20 cached, a 1,050,000-token context, 128,000 output tokens and an April 2026 knowledge cutoff.

The cheapest GPT-6: $0.10 in and $0.50 out per million tokens, $0.01 cached, with the same 1,050,000-token context and 128,000-token output limit as GPT-6 Sol and a May 2026 knowledge cutoff.

Anthropic’s new default model at $4 in and $20 out per million tokens, a 1M-token context, 128,000 output tokens, a June 2026 cutoff — and a published retirement floor of 22 September 2027, a year to the day from release.

A Postgres-compatible engine that puts full-text, vector and hybrid search in the same database as columnar analytics; the single-node build is Apache 2.0 and the first production release, Krummelanke, shipped on 22 September 2026.

Anomalo’s agent layer for data teams: it ranks every change it finds in a warehouse table by magnitude, investigates the ones that matter and writes the finding up as a report, without anyone writing SQL or opening a dashboard.

A Model Context Protocol server that lets an AI assistant make and edit product videos in a Clueso account — walkthroughs, release notes, training and onboarding clips — by conversation rather than through the editor.

Documents, slides, forms and sheets that answer back: upload a spreadsheet, deck or PDF and WZRD turns it into a presentation that can talk an audience through the material rather than simply display it.

Real-user monitoring bolted onto an uptime checker: one async script records LCP, CLS, INP and FCP from actual visitors, alongside DNS, connection, time to first byte and total load time on every page load.

An open-weight 6B text-to-image model from Ant Group’s inclusionAI lab for interface screens, dashboards, infographics and posters with legible text and transparent backgrounds, under the MIT licence.

An open-weight 6B model from Ant Group’s inclusionAI that splits a finished, flattened design back into separate transparent layers that can be edited, under the MIT licence.

A crew of AI agents that turns a written design brief into researched design directions, production HTML with Tailwind and a design system. Free in alpha; you bring your own model key and pay the model provider.

A writing coach from Swat.io that helps employees write their own LinkedIn posts through guided questions rather than generating them, with team campaigns and a shared calendar. Free for up to five people.

BottleCap AI’s fine-tune of Qwen3.8-27B that thinks less: 37% fewer reasoning tokens for under one point of accuracy on the developer’s tests. Free for small businesses; larger ones need a commercial licence.

One interface through which AI agents reach official data providers, Firecrawl’s own research, developer and government indexes, and the live web — with providers paid when agents use their data. Runs on Firecrawl credits.