October 2026.1
This week the focus is on the mid tier and on European open weights: Sonnet 5.5 and GPT-6.1 Sol both land at $2/$10 per million tokens while getting close to their flagships, and Aleph Alpha releases Kolibri, an Apache-licensed English-German model built for on-premise use.
📖 Story 1: Kolibri: A Sovereign Open-Weight Model
aleph-alpha.com · Read
Aleph Alpha has released Kolibri, an English-German open-weight model, timed for the Day of German Reunification.
It is a Mixture-of-Experts model with 78B total parameters and about 3.5B active per token, a context window of up to 1M tokens, and Apache 2.0 weights on Hugging Face. Roughly 21% of the 20T pre-training tokens are German, mostly organic web text and LLM rephrasings of German documents rather than translations from English. A custom tokenizer splits German compounds along their natural parts, so the same text needs fewer tokens.
The company positions it for regulated work in public administration, industry and aerospace: small enough to run on-premise, trained in Germany and Finland, and tuned to say "I don't know" when the answer is not in the provided documents. On its own benchmarks it keeps up with Qwen3.6 35B-A3B and Nemotron 3 Super, but the dense Qwen3.8 27B still scores higher overall, including in German (79.9 vs 70.8).
Kolibri is built to keep regulated customers' data in-house, not to top leaderboards.
💬 HN Discussion
The Hacker News discussion split between respect for the openness and doubt about the model itself.
The most repeated criticism was that Qwen3.8 27B beats Kolibri on nearly every benchmark in Aleph Alpha's own table. One team member replied that a dense 27B model and a sparse MoE sit at different points of the cost trade-off: Kolibri needs more memory but less compute per token.
Others questioned the sovereignty label given the pending Cohere merger. Early testers reported mixed results, including a confident answer about a made-up song and an outdated German legal reference.
Still, many praised the tech report as unusually open and detailed. For now, transparency is its strongest asset.
📖 Story 2: Claude Sonnet 5.5
anthropic.com · Read
Anthropic has released Claude Sonnet 5.5, the mid-tier model in its 5.5 family, positioned as a faster and cheaper complement to Opus 5.5.
Pricing is unchanged from Sonnet 5 at $2 per million input tokens and $10 per million output tokens, half the Opus 5.5 rates. Anthropic says it needs far fewer tokens per task, cutting cost by up to 30% and running more than 30% faster.
The benchmark jump is large. On Terminal-Bench 4.0 it scores 70.6%, up from 10.3% for Sonnet 5 and above Opus 5.5's 66.4%. On most other coding and knowledge-work tests it lands within two or three points of Opus. Anthropic itself says Opus remains clearly stronger on open-ended work that needs sustained judgment. Customer reports back the efficiency claim: one finance firm measured about 121k tokens per answer against 497k for Sonnet 5.
It is also the first Sonnet launched with cyber safeguards, so higher-risk security requests fall back to Sonnet 5.
The story is less a smarter model than near-Opus capability at a lower price per task.
💬 HN Discussion
The Hacker News discussion was impressed by the benchmarks but quickly turned to the cyber safeguards.
Several developers reported being downgraded to older models for ordinary work such as verifying a write-ahead log, reviewing old C code or Bluetooth presence detection. Members of the Cyber Verification Program said it made no difference, though others noted the program does not yet cover the 5.5 models. Some moved those tasks to cheap open-weight models.
A second thread asked where Sonnet fits at all. Some Max-plan users said Opus 5.5 is so efficient they rarely hit limits; others noted cache reads cost the same on both. Replies suggested per-token API work or subagents.
The model looks strong; the guardrails may push people elsewhere.
📖 Story 3: GPT 6.1 Sol: Near-Astra intelligence for a fifth of the price
openai.com · Read
OpenAI has released GPT-6.1 Sol, an update to its mid-tier model that it says nearly matches GPT-6 Astra on agentic coding, computer use and professional work at a fifth of the price.
API pricing is $2 per million input tokens, $10 per million output tokens and $0.10 per million cached input tokens, against $10, $50 and $1 for Astra. The cached price, half that of GPT-6 Sol, matters most for agents that keep resending the same context.
OpenAI's own charts claim Sol matches Astra on the DeepSWE coding benchmark and edges past Opus 5.5 on business-workflow and PDF tasks at a fraction of the cost. On a scientific-workflow benchmark it averages $5.47 per task against roughly $23 for Opus 5.5 and Astra, though Astra still scores highest. Factual errors on hard prompts fell from 11.4% to 7.7% at low reasoning effort.
The model is available in ChatGPT Work, Codex and the API, but not yet in regular chat.
If OpenAI's numbers hold, Sol becomes the default for everyday agent work, with Astra kept for the hardest scientific tasks.
💬 HN Discussion
The Hacker News discussion read the release mostly as a correction. Many commenters had found GPT-6 Sol worse than the 5.6 generation and saw 6.1 as a rushed answer to Opus 5.5.
The bigger argument was about what price cuts signal. Some saw capability gains flattening and competition shifting to cost, pushed by cheap Chinese open-weight models. Others pointed to clear recent progress in coding, maths and 3D work.
One hands-on test stood out: on the same task, Astra burned more than two weeks of subscription budget while Opus used a fifth of one and got further.
Benchmarks set the headline, but cost per finished task is what users now measure.
💬 Community Moment
Leaked image of hardware running 6.1 Sol
https://www.reddit.com/r/codex/comments/1wwls9e/leaked_image_of_hardware_running_61_sol/🛠️ Projects Worth Checking Out
- GitHub - contemper-project/contemper: Build bootable VM images from container images.
- GitHub - andreglegg/forge: Local-first coding agent CLI optimized for local and small language models.
- GitHub - datasette/datasette-mcp: Adds a /-/mcp MCP server to any Datasette instance
- GitHub - block/buzz: A hive mind communication platform
- GitHub - postgresql-tools/wal-g: PostgreSQL WAL archiving, backups, and disaster recovery. Actively maintained successor to WAL-G.
