https://thenexttoken.net An independent publication about artificial intelligence: the news that matters, how the technology works, the research behind it and the products worth using. en https://thenexttoken.net/posts/coding-benchmark-answer-leaks https://thenexttoken.net/posts/coding-benchmark-answer-leaks Thu, 22 Oct 2026 00:00:00 GMT Research Agents can read the fix from Git history, PyPI or GitHub, or edit the tests. Which coding benchmarks close each leak, and what closing them does to scores. https://thenexttoken.net/posts/eu-ai-act-ai-content-marking https://thenexttoken.net/posts/eu-ai-act-ai-content-marking Thu, 22 Oct 2026 00:00:00 GMT News If your app generates text, images or audio for people in the EU, Article 50 marking is likely your duty. Who it falls on, what to mark and from when. https://thenexttoken.net/posts/cheap-models-for-data-extraction https://thenexttoken.net/posts/cheap-models-for-data-extraction Wed, 21 Oct 2026 00:00:00 GMT Products On 104 synthetic documents, GPT-6 Luna and Claude Haiku 5.5 got nearly every record right for under $0.25 per 1,000. Valid JSON never meant correct fields. https://thenexttoken.net/posts/how-to-grade-an-ai-claim https://thenexttoken.net/posts/how-to-grade-an-ai-claim Wed, 21 Oct 2026 00:00:00 GMT Notes Grade an AI claim by what its evidence has done: rumoured, announced, documented, available, measured or reproduced. Worked on five October 2026 claims. https://thenexttoken.net/posts/max-tokens-rate-limits https://thenexttoken.net/posts/max-tokens-rate-limits Wed, 14 Oct 2026 00:00:00 GMT How it works Anthropic's API has ignored max_tokens for rate limits since February 2026; OpenAI, Azure, Bedrock and Databricks charge it before a request runs. https://thenexttoken.net/posts/why-temperature-zero-is-not-deterministic https://thenexttoken.net/posts/why-temperature-zero-is-not-deterministic Tue, 13 Oct 2026 00:00:00 GMT How it works Temperature 0 always picks the top token, but server batching changes the arithmetic behind the scores. Why outputs vary, and what seeds really promise. https://thenexttoken.net/posts/why-token-counts-differ https://thenexttoken.net/posts/why-token-counts-differ Tue, 13 Oct 2026 00:00:00 GMT How it works Each model family splits text with its own tokeniser, so one prompt can need 5% more tokens or four times as many. Measured by content type and language. https://thenexttoken.net/posts/how-much-vram-a-local-model-needs https://thenexttoken.net/posts/how-much-vram-a-local-model-needs Mon, 12 Oct 2026 00:00:00 GMT How it works A local model needs its weights plus a KV cache sized for the whole context window. How to work out both from config.json, and why runtimes fill the card. https://thenexttoken.net/posts/how-ai-text-watermarks-work https://thenexttoken.net/posts/how-ai-text-watermarks-work Mon, 28 Sep 2026 00:00:00 GMT How it works Text watermarks bias word choice with a secret key. How detection works, what a positive result proves, and why edits and paraphrase weaken the mark. https://thenexttoken.net/posts/how-prompt-caching-works https://thenexttoken.net/posts/how-prompt-caching-works Sun, 27 Sep 2026 00:00:00 GMT How it works A prompt cache reuses the model's computed state for an identical prefix, so one early edit can re-bill the whole request. How it works and what it costs. https://thenexttoken.net/posts/what-reasoning-effort-controls https://thenexttoken.net/posts/what-reasoning-effort-controls Sun, 27 Sep 2026 00:00:00 GMT How it works Reasoning effort on Anthropic, OpenAI and Gemini APIs guides how much work a model does; it is not a token count. What it changes, and when more effort pays. https://thenexttoken.net/posts/why-new-model-apis-reject-old-parameters https://thenexttoken.net/posts/why-new-model-apis-reject-old-parameters Sun, 27 Sep 2026 00:00:00 GMT How it works New Claude, GPT-6 and Gemini 3.8 Flash models reject or ignore old request patterns. What replaces each, and which reasons are documented.