A week of new models
New frontier models and a security breach as April draws to a close.
Anthropic, OpenAI, and DeepSeek each made frontier-model news between Tuesday and Friday last week. Anthropic confirmed on Tuesday that an unauthorized group had accessed its restricted Mythos Preview through a contractor’s credentials. OpenAI released GPT-5.5 on Thursday, breaking a three-way tie with Anthropic and Google at the top of the leaderboard. DeepSeek released the long-anticipated V4 in preview on Friday, the company’s first frontier-class model trained on Chinese silicon.
Top of the pile
OpenAI released GPT-5.5 last Thursday, six weeks after GPT-5.4 and one week after Anthropic’s Opus 4.7. GPT-5.5 is the company’s first ground-up retrained base model since GPT-4.5. Each intervening 5.x version layered post-training on the same base. The company is positioning the new model as agent-first, describing it as a system that takes actions, uses tools, checks its work, and continues through ambiguity without step-by-step prompting.
On the Artificial Analysis Intelligence Index, GPT-5.5 scored 60, three points ahead of Claude Opus 4.7 and Gemini 3.1 Pro Preview, breaking a three-way tie at the top. The composite collapses a more mixed picture into a single number, with GPT-5.5 placing second to Gemini 3.1 Pro Preview on three of the ten evaluations that compose this index. Per-token API pricing doubled, with GPT-5.5 charging $5 per million input tokens and $30 per million output tokens, up from GPT-5.4’s $2.50 and $15. OpenAI defends the increase by pointing to token efficiency. According to Artificial Analysis, GPT-5.5 generates roughly 40 percent fewer tokens than 5.4 on the same Intelligence Index run, leaving net costs around 20 percent higher.
The open-source tiger
DeepSeek released V4 in preview last Friday, open-sourcing both V4-Pro and V4-Flash on Hugging Face. V4-Pro carries 1.6 trillion parameters and a 1-million-token context window. DeepSeek’s technical report claims that V4-Pro outperforms all other open-source models in agentic coding and reasoning. The same report says that V4-Pro “falls marginally short of GPT-5.4 and Gemini 3.1 Pro,” characterizing the gap as a trajectory trailing frontier models by three to six months.
DeepSeek trained V4 on Huawei Ascend and Cambricon chips, with Huawei publicly confirming on launch day that its Ascend 950 clusters had supported the model. R1, the model that put DeepSeek on the map last year, ran on NVIDIA hardware. V4 is the company’s first frontier-class release on primarily domestic Chinese silicon. The day before, White House OSTP director Michael Kratsios issued a memo accusing foreign entities, primarily in China, of running “industrial-scale” campaigns to extract capabilities from US frontier AI models through distillation.
Mythos draws fire
Anthropic announced Claude Mythos Preview on April 7 under a restricted-access program called Project Glasswing, distributing the model to roughly forty vetted partners. Last Tuesday, Anthropic confirmed that it was investigating reports of unauthorized access through a third-party vendor environment. The access began on the day of the announcement. One member of a small private group reportedly used contractor credentials to reach the vendor environment, then drew on previously leaked information about Anthropic’s infrastructure to locate the model. Contrast Security CISO David Lindner told Fortune that with thousands of users across partner companies, the leak was “bound to happen.”