The week in AI

OpenAI had two announcements this week. On 22 June it launched Daybreak, a cyber defense program built around a full version of GPT-5.5-Cyber. Two days later it unveiled the Jalapeno chip. Anthropic put Claude into Slack channels as a shared teammate with Claude Tag on 23 June. DeepSeek released DSpark on 27 June, a speculative decoding method that it says makes V4 generate text much faster for each user.

ByteDance Seed shipped Seed 2.1, Alibaba's Qwen team released world models that simulate agent environments, and Mistral released a new OCR model.

Jalapeno is built for inference, the work of running trained models for users. It is a new design for LLM serving, and OpenAI and Broadcom did not adapt it from a general-purpose accelerator. OpenAI says the chip went from design to tape-out in nine months, which it calls the fastest ASIC cycle. Tape-out is the point where a finished design goes to the fab.

OpenAI also says its own models helped design the chip. The announcement did not say which models did which parts of the work.

The 24 June announcement included no performance results. It also gave no details on which workloads or models Jalapeno would serve first.

Daybreak packages several things. Vetted defenders get trusted access to cyber models, and the program centres on the full version of GPT-5.5-Cyber. Codex Security workflows help teams triage findings and fix them. An open-source effort called "Patch the Planet" applies the same tools to shared code. Access is by application, and OpenAI lists pricing in its API changelog.

OpenAI reports that Codex Security's cloud service has scanned more than 30 million commits across more than 30,000 codebases since March. OpenAI also says humans have marked more than 70,000 of its findings as fixed. These figures come from OpenAI and are not independently measured.

OpenAI is aiming the program at the gap between finding a bug and shipping a fix, because defenders tend to get stuck at that step.

Each channel gets a single Claude that everyone in the channel can see. Claude Tag keeps context across conversations and can run asynchronous tasks that last hours or days. Teams can also turn on an optional "ambient" mode, where Claude starts work without being asked. Admins decide which tools, data and channels Claude can use.

Anthropic says an internal version of Claude Tag now creates 65% of the code on its own product team. Anthropic gave that figure at launch and did not say how it measured it.

Claude Tag is in beta for Enterprise and Team plans.

Speculative decoding uses a small drafter model to guess the next several tokens, and the large model then checks the guesses in one pass. DSpark's drafter is semi-autoregressive, so it predicts tokens in blocks and not one at a time, using three blocks. It also decides how much drafted text to verify based on how confident the drafter is and how loaded the server is. MTP-1 is the baseline, where the model predicts one extra token at each step.

DeepSeek published DSpark modules for V4-Pro and V4-Flash. It also released DeepSpec, a codebase for training and evaluating drafters. DeepSpec covers DSpark, DFlash and EAGLE3, and comes with drafters for Qwen3 and Gemma-4.

The speedup was measured by DeepSeek inside its own serving system. Results on other serving stacks have not been reported.

• Seed 2.1 from ByteDance Seed shipped on 23 June in Pro and Turbo sizes for office and coding agent work, and ByteDance reports 53.0 on Workspace Bench and 47.0 on NL2Repo-Bench for Pro, against 58.2 for Claude Opus 4.7 on NL2Repo-Bench.
• ByteDance also says the reinforcement learning (RL) training for Seed 2.1 cut the average number of steps in graphical user interface (GUI) tasks by 16%, and it claims the highest GDPVal score and the best MobileWorld result.
• Qwen-AgentWorld from Alibaba's Qwen team arrived on 22 June as language world models in 35B-A3B and 397B-A17B sizes, which predict an agent environment's next state and were trained on more than 10 million interaction trajectories across seven domains.
• Qwen trained AgentWorld in three stages, continued pretraining, supervised fine-tuning (SFT), and RL with rubric and rule rewards, but only the 35B-A3B checkpoint is on Hugging Face.
• Mistral OCR 4 came out on 23 June and returns bounding boxes, block labels and confidence scores in 170 languages, and Mistral reports a 72% average win rate in human preference tests.