The week in AI

GPT-5.4 was the week's main release. It folds OpenAI's Codex coding line into the flagship model. On 6 March, Anthropic published two reports on Claude Opus 4.6. In the first, the model found 22 Firefox vulnerabilities in two weeks of joint work with Mozilla. The second described how the model twice worked out that it was being tested on BrowseComp and decrypted the answer key.

OpenAI also published research on 5 March showing that reasoning models rarely manage to shape their chains of thought, even when told they are being monitored. In people news, Lin Junyang stepped down as technical lead of Alibaba's Qwen team.

OSWorld-Verified measures whether a model can complete tasks on a desktop computer by operating the screen, keyboard and mouse. OpenAI cites a human baseline of 72.4% on the same benchmark, so its reported score puts the model slightly above that line. GPT-5.4 has a 1 million token context window. In the API it adds tool search, which lets the model find the tools it needs for a task, and compaction, which shortens long working histories so sessions can run longer.

The release merges the Codex coding models into the general flagship. OpenAI reports 57.7% on the public SWE-Bench Pro set, against 56.8% for GPT-5.3-Codex. That means the merged model roughly matches the dedicated coding model.

OpenAI reports 83.0% wins or ties on GDPval, up from 70.9% for GPT-5.2. GDPval has graders compare model output on professional work tasks with output from human professionals. OpenAI also reports 33% fewer claim-level factual errors than GPT-5.2 and 18% fewer false responses, both measured on a set of prompts that users had flagged. The company names spreadsheets, slides and deep research as areas of improvement.

GPT-5.4 costs $2.50 per million input tokens and $15 per million output tokens through a closed API. Two other OpenAI releases landed in the same week. GPT-5.3 Instant, out on 3 March, updated ChatGPT's default fast model. Codex Security entered research preview on 6 March.

The work was a joint project with Mozilla, and Mozilla assigned the severity ratings. According to Anthropic, the 14 high-severity bugs equal almost a fifth of all high-severity Firefox bugs remediated in 2025.

Anthropic also tried to get the model to turn the bugs into working exploits. Across several hundred attempts, costing about $4,000 in API credits, it succeeded twice. Both exploits were crude, and both worked only in a test environment with Firefox's sandbox disabled.

For this model, then, finding bugs was much cheaper than exploiting them. OpenAI's Codex Security, released on the same day, targets the same kind of work. It is an application security agent, previously called Aardvark, that checks its own findings and proposes fixes.

BrowseComp tests whether a model can find hard-to-locate facts by browsing the web. Anthropic ran Opus 4.6 in a multi-agent setup across 1,266 problems and found leaked answers on 11 of them.

Nine of those 11 were ordinary contamination. The answers had been published online, for example in ICLR 2026 submissions, and the model found them during its search. The other two showed a pattern Anthropic had not described before. The model suspected it was in an evaluation, worked out which benchmark it was, and then located and decrypted the answer key.

Anthropic says the result raises doubts about static benchmarks when models have web access. Two problems out of 1,266 barely moves the score. The concern is that the model went looking for the answer key unprompted.

The paper introduces a measure called CoT controllability. It asks whether a model can deliberately shape its visible reasoning trace, for example to hide its intentions from a monitor reading that trace. Across the frontier models OpenAI tested, controllability was low. It was somewhat higher on simple tasks.

OpenAI presents this as support for chain of thought monitoring as a safety method, because a model that cannot easily edit its reasoning is harder for it to disguise. The paper is a research preview. It does not settle whether the result will hold as models get more capable.

• Phi-4-reasoning-vision-15B is an open 15 billion parameter vision-language model from Microsoft, released on 4 March, that decides when to reason, and Microsoft reports 75.2% on MathVista and 88.2% on ScreenSpot v2 after training on only 200 billion multimodal tokens.
• Gemini 3.1 Flash-Lite, released by Google DeepMind on 3 March, is the first Flash-Lite model in the Gemini 3 series, priced at $0.25 per million input tokens and $1.50 per million output tokens, and Google reports 2.5 times faster time to first token than Gemini 2.5 Flash.
• GPT-5.3 Instant, released on 3 March, updates ChatGPT's default fast model with fewer needless refusals and caveats and better grounded web answers.

• Lin Junyang, technical lead of Alibaba's Qwen models, resigned on 3 March while Tongyi Lab was reorganizing its model teams, and other Qwen members were reported to have resigned the same week.
• Caitlin Kalinowski, head of hardware and robotics at OpenAI, resigned on 7 March over OpenAI's Pentagon deal, saying surveillance and lethal autonomy deserved more deliberation.