OpenAI just launched fine-tuning for GPT-4o across all paid developer tiers, giving teams a way to tailor its flagship model with as few as a dozen custom examples. Training costs $25 per million tokens, but organizations get 1M free training tokens daily through September 23.
Perplexity just started rolling out a built-in code interpreter with live plot rendering. The tool can install third-party Python libraries on the fly and display charts directly in answers—like pulling stock data with yfinance. It is currently rolling out gradually to users.
Microsoft is working on a new Workspaces feature for its PowerToys suite on Windows. The tool lets users save multi-app layouts to a shortcut, automatically launching and snapping apps across displays with a single click. It is slated to debut in the PowerToys 0.84 update.
Anthropic just rolled out prompt caching on the Anthropic API for Claude 3.5 Sonnet and other models. By storing frequently used context between calls, developers can slash input costs by up to 90% and cut latency by up to 85% on long prompts. It is a massive win for coding assistants, deep agent loops, and chatting with whole books.
OpenAI just took back the #1 spot on the LMSYS Chatbot Arena leaderboard with its updated ChatGPT-4o model, edging past Google's Gemini-1.5-Pro-Exp with a 1314 score. Tested anonymously across over 11,000 community votes, the new checkpoint shows noticeable gains in coding (up 30+ points) and hard prompts, and developers can already call it in the API.
OpenAI just launched SWE-bench Verified, a human-validated subset of the popular software engineering benchmark built with its original authors. Working with 93 professional Python developers, the team filtered out ambiguous issue descriptions and unfair unit tests to curate 500 reliable tasks. Why it matters: Coding benchmarks often penalize correct code with broken unit tests, and this cleaner test set gives the industry a far more accurate gauge of real-world agent capabilities.
Cosine just launched Genie, an autonomous AI software engineer that scored 30% on the SWE-Bench benchmark, more than doubling Devin's 13.8%. Built on a fine-tuned, long-output variant of OpenAI's GPT-4o, Genie is trained to emulate human reasoning across 15 programming languages. Why it matters: The race for an autonomous artificial colleague just got a lot faster.
Google just slashed Gemini 1.5 Flash pricing, cutting input costs by 78% to $0.075 per million tokens and output by 71% starting August 12. The update also rolls out text tuning to all developers via AI Studio and the Gemini API, alongside native multimodal understanding for image-heavy PDFs.
OpenAI just rolled out Structured Outputs in its API, guaranteeing that model responses strictly follow developer-provided JSON Schemas with a perfect 100% evaluation score on the new gpt-4o checkpoint. Using constrained decoding, it eliminates messy prompt hacking and retry loops for developers. Why it matters: building reliable agents and data pipelines just became significantly less headache-inducing.
OpenAI just confirmed that its upcoming DevDay won't feature its next major flagship model. Instead of a splashy keynote, the company is shifting to a three-city roadshow across San Francisco, London, and Singapore to focus primarily on API updates and developer sessions.
Mistral AI just dropped Mistral Large 2, a 123B-parameter flagship model built with a 128k context window designed to run efficiently on a single node. It delivers major gains in coding and math that rival GPT-4o and Claude 3 Opus, while actively curbing hallucinations. Non-commercial weights are free to download, offering serious frontier power without giant infrastructure.
Cohere just launched Rerank 3 Nimble, a new foundation model built for enterprise search and RAG workflows. It delivers roughly 3x faster performance than Rerank 3 while keeping high retrieval accuracy, but it is currently exclusive to Amazon SageMaker.
Meta just dropped its open-source Llama 3.1 family, headlined by a massive 405B model that rivals GPT-4o and Claude 3.5 Sonnet across key benchmarks. The entire lineup supports a 128K context window, and Meta has updated its license so developers can finally use 405B outputs to distill smaller models.
Mistral AI just launched Codestral Mamba, a code-focused model built with Albert Gu and Tri Dao that swaps Transformers for the Mamba architecture. It delivers linear-time inference and handles tested in-context retrieval up to 256k tokens without slowdowns. Even better, Mistral open-sourced it under Apache 2.0 for free local deployment.
Anthropic just updated Claude so users can publish and share their Artifacts, as well as remix creations made by others. The sidebar feature lets you render code, interactive diagrams, and mini-games right next to chat, turning AI generations into easily shareable mini-apps.
Hugging Face just brought Google Cloud TPU v5e chips to its Inference Endpoints and Spaces, starting at $1.38 an hour. Powered by the open-source Optimum TPU library, developers can deploy and serve open models like Gemma, Llama, and Mistral in a few clicks. It gives builders a viable, budget-friendly alternative to scarce Nvidia GPUs.
Anthropic just launched an Evaluate tool directly inside the Anthropic Console to streamline prompt tuning. Powered by Claude 3.5 Sonnet, it can automatically draft prompts and realistic test cases, import datasets from CSVs, and let teams compare outputs side-by-side with a 5-point grading scale. Basically, prompt engineering just got treated like real software testing.
Poe just introduced Previews, letting users see and interact with web applications built directly inside chat threads. The feature works alongside coding powerhouses like Claude 3.5 Sonnet, GPT-4o, and Gemini 1.5 Pro. Basically, you can now generate frontend projects and test the live widgets immediately without pasting code elsewhere.
LangChain just rolled out the stable release of LangGraph v0.1 alongside a closed beta of LangGraph Cloud. The framework ditches rigid abstractions to give developers low-level control over multi-agent workflows, complete with human-in-the-loop approvals, time-travel debugging, and dedicated infrastructure for scaling agent fleets reliably.
OpenAI just introduced CriticGPT, a model based on GPT-4 designed to critique ChatGPT's code and help human trainers spot bugs. In experiments, trainers assisted by CriticGPT outperformed unassisted reviewers 60% of the time. The lab is now integrating these critic models into its RLHF pipeline to keep superhuman AI properly aligned.
