Language Model News

2026/10/13

  1. How much does speech-to-text cost per hour? Comparing the xAI and Mistral APIs
    PricexAIMistral
  2. Mistral Large 4 costs $1.36 in and $4.18 out, at half price for two weeks after launch
    PriceMistral
  3. Claude Managed Agents' web_fetch now fetches only URLs that already appear in the session
    How to UseAnthropic
  4. xAI's speech-to-text API costs $0.10 an hour in batch, and moved to 2.0 on 2 October
    PricexAI
  5. Claude Max and Team plans now include monthly API credits ($100 on Max 5x)
    PriceAnthropic
  6. GPT-6.1 Sol charges less than GPT-6 Sol for cached input
    ModelOpenAI
  7. Claude Sonnet 5.5 keeps Sonnet 5's price, and its cache reads were halved on 7 October
    ModelAnthropic
  8. Claude Haiku 5.5 is priced like GPT-6 Luna up to 100K tokens, and five times higher beyond
    PriceAnthropicOpenAI

2026/09/29

  1. Gemini 2.5 models are now limited to users who have used them before
    ModelGoogle
  2. Grok 4.7 joins xAI's model list at the same price and context as Grok 4.6
    ModelxAI
  3. Gemini 3.8 Live pricing: audio in at $0.005 a minute and audio out at $0.018
    PriceGoogle

2026/09/28

  1. Gemini 3.8 Flash TTS and Flash-Lite TTS prices double on 1 January 2027
    PriceGoogle
  2. Claude Opus 5.5 is $4 in and $20 out, and the model Anthropic now recommends first
    ModelAnthropic
  3. GPT-6 Sol and GPT-6 Luna: Luna costs a twentieth of Sol
    ModelOpenAI

2026/09/15

  1. DeepSeek V4.1 Flash takes over from V4 Flash at a lower price
    ModelDeepSeek
  2. Sakana AI's Fugu Ultra v2 and Fugu Max answer by routing work to other models
    ModelSakana AIOpenAI
  3. Cognition's SWE-2 is a coding model built on Kimi K3
    ModelCognitionMoonshot
  4. Inception's Mercury 2.5, a diffusion model at $0.75 per million output tokens
    PriceInceptionOpenAIAnthropic
  5. GPT-Live 1 bills by the minute, at $0.05, with the reasoning model on top
    PriceOpenAI

2026/09/08

  1. Meta sells the same model at two prices, and the cheap one is paid for with your prompts
    PriceMeta
  2. Grok 4.6 is the flagship, and it holds half the context of Grok 4.3
    ModelxAI
  3. Kimi K3 holds a million tokens at one flat rate
    ModelMoonshot
  4. The same Qwen model costs less from Beijing than from Singapore
    PriceAlibaba
  5. On Mistral, Large costs a third of Medium
    PriceMistral
  6. GPT-6 Astra and the three models that make up GPT-6
    ModelOpenAI
  7. GPT-6 Astra prices, and the 272K line that reprices the whole request
    PriceOpenAIAnthropic
  8. Claude Fable 5.1 takes the top of the lineup, and Fable 5 moves to legacy
    ModelAnthropic
  9. Cache reads on Fable 5.1 are a quarter of what they cost on Fable 5
    CostAnthropic

2026/09/02

  1. The current Claude lineup: 1M context, 128K output, retirement more than a year out
    ModelAnthropic
  2. SWE-bench Verified is at 97, and the top of the table is hard to read
    BenchmarkAnthropic
  3. A benchmark number means nothing without the conditions it was taken under
    BenchmarkAnthropicOpenAIGoogleDeepSeek
  4. Claude API prices, and the Sonnet 5 increase that did not happen
    PriceAnthropic
  5. OpenAI API prices: the GPT-5.6 line has long-context rates too
    PriceOpenAI
  6. Gemini API prices, and the two dates that change them
    PriceGoogle
  7. DeepSeek now charges double during peak hours
    PriceDeepSeek
  8. Perplexity Sonar ends on 27 September 2026, replaced by the Agent API
    PricePerplexity
  9. What prompt caching actually costs, and where the tenth does not hold
    CostOpenAIAnthropicGoogleDeepSeekxAIMetaMistralAlibabaMoonshotSakana AI
  10. Long context: who reprices the whole request, and who does not
    CostOpenAIAnthropicGooglexAIAlibabaMoonshotSakana AI
  11. Adding an MCP server to Claude Code, and which scope to pick
    How to UseAnthropic

SWE-bench Verified is at 97, and the top of the table is hard to read

BenchmarkAnthropic

SWE-bench Verified measures whether a patch written by a model makes a real repository test suite pass⁠[1]. The score comes from running those tests in a container, not from preference judgements or synthetic checks⁠[1].

On the Steel leaderboard, Claude Opus 5 leads at 97.0 percent from an independent Vals.ai run (September 2026), with a second Claude Opus 5 entry at 96.0 (July 2026)⁠[1]. Then come Claude Mythos 5 at 95.5 (June 2026), Claude Fable 5 at 95.0 (June 2026) and Claude Mythos Preview at 93.9 (April 2026)⁠[1].

Conditions differ per submission, and the two Opus 5 rows show how much. The 97.0 came from an independent run on a bash-only harness in isolated Docker containers⁠[1]. The 96.0 is a five-trial average in the standard configuration⁠[1]. The 95.5 on Mythos 5 is an average over five trials with adaptive thinking at maximum effort⁠[1]. Set beside a number from a single trial, none of these are measuring the same thing.

The leaderboard attaches its own warning. It calls the benchmark mature and heavily exposed in public training data, and says top frontier scores should be read with contamination and test-design caveats, especially when comparing very high-scoring systems⁠[1]. It calls the benchmark strong for public issue-resolution work and weaker as a frontier signal once scores approach saturation⁠[1].

The official SWE-bench site lists five variants: the original, Verified, Multilingual, Multimodal and Lite⁠[2]. A score is comparable only against another score on the same variant.

Sources

  1. Steel - SWE-bench Verified leaderboard
  2. SWE-bench