AI model news, week of 21 September 2026
In the week of 21 September 2026, Anthropic, Google, OpenAI, Together AI, DeepSeek, Mistral AI, and xAI published 11 updates that matter for developers: new models (6), price changes (1), deprecations and retirements (1), and announcements (3).
New models
- Together AI launches together/Tev1-4B-experimental
Together AI released together/Tev1-4B-experimental, a classification model built on top of Qwen3.5 4B and available on its serverless platform.
- Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS are generally available
Google released two text-to-speech models in the Gemini API as GA, plus a new Voices endpoint at /v1beta/voices.
- Grok 4.7 launches on the xAI API
xAI released grok-4.7 on its API with a 500k context window and different prices below and above 200k prompt tokens.
- DeepSeek-V4.1-Flash release with native multimodal support
DeepSeek released V4.1-Flash on its API, with native multimodal support and the model name deepseek-flash.
- OpenAI releases GPT-6 Sol and GPT-6 Luna in the API
OpenAI released GPT-6 Sol (gpt-6-sol) and GPT-6 Luna (gpt-6-luna), two reasoning models that accept text and image inputs.
- Claude Opus 5.5 launches at $4 / $20 USD per MTok
Claude Opus 5.5 is now in the API with a 1M token context window, 128k max output and adaptive thinking that is always on.
Price changes
- Anthropic resumes billing for specific API refusals
Anthropic now bills for API requests that are refused before any output when the refusal category is bio, frontier_llm, or reasoning_extraction.
Deprecations and retirements
- Together AI deprecates four models on serverless
Together AI has deprecated openai/gpt-oss-20b, google/gemma-4-31B-it, thinkingmachines/Inkling-Small, and intfloat/multilingual-e5-large-instruct on serverless.
Announcements
- OpenAI fixes image encoding bug in GPT-6 Sol and GPT-6 Luna
The update improves results on visual tasks in the API and Codex, including computer use.
- OCR 4.1 from Mistral is now generally available
Mistral has made the OCR 4.1 model generally available.
- Google restricts Gemini 2.5 API access to existing active users
Google is limiting access to Gemini 2.5 models to past active users and directs new projects to 3.5 Flash-Lite or 3.8 Flash.
Models from this week on Zabio
| Model | Input, $/M | Output, $/M | Cached, $/M | Context window | 1,000 calls | Writing, medium |
|---|---|---|---|---|---|---|
gemini-3.8-flash-ttsGoogle | 0.50 | 0.00 | no data | 0 | $1.00 | no data |
gemini-3.8-flash-lite-ttsGoogle | 0.50 | 0.00 | no data | 0 | $1.00 | no data |
grok-4.7xAI | 2.20 | 6.60 | 0.55 | 500,000 | $7.70 | 79.1 |
deepseek-flashDeepSeek | 0.30 | 1.20 | 0.006 | 1,048,576 | $1.20 | 76.0 |
deepseek-v4-flashDeepSeek | 0.20 | 0.40 | 0.04 | 1,000,000 | $0.60 | 66.3 |
deepseek-v4-flash-vision-expDeepSeek | 0.2156 | 0.6468 | 0.00686 | 131,072 | $0.75 | 76.6 |
deepseek-v4-proDeepSeek | 1.32 | 3.96 | 0.044 | 1,000,000 | $4.62 | 67.2 |
gpt-6-solOpenAI | 2.00 | 10.00 | 0.20 | 400,000 | $9.00 | 77.7 |
gpt-6-lunaOpenAI | 0.10 | 0.50 | 0.01 | 400,000 | $0.45 | 61.8 |
claude-opus-5-5Anthropic | 4.00 | 20.00 | 0.20 | 1,000,000 | $18.00 | 75.3 |
claude-opus-5Anthropic | 5.00 | 25.00 | 0.50 | 1,000,000 | $22.50 | 77.2 |
Prices are the vendor list prices in USD per million tokens. "1,000 calls" is 1,000 requests of 2,000 input tokens and 500 output tokens each, without caching. Zabio bills the list price plus its service fee: see pricing.
Scores: percent correct for that kind of job, aggregated from public results by LMArena and LiveBench, read on 16 November 2026. LMArena: CC BY 4.0, LiveBench: Leaderboard results used with attribution; the 2.0 release states no terms of its own (read 2026-11-06); the first version's Mathematics, Language, IF results: Code Apache 2.0 and MIT; leaderboard results used with attribution