CodeMender expands security review with isolation and human control
Google adds scanning and isolation options to its code-security agent, which remains in preview.
AI models for business
Choose for the work
you need to complete.
A model may excel at one task and be unsuitable for another. Explore announcements, limits and comparison criteria using examples from your business.
A useful comparison starts with a task.
Quality
Check accuracy, sources, format and necessary corrections. Include inputs in the languages your team uses.
Time and cost
Measure the complete task, including retries, tools and human review.
Tools
Check the connections, files and actions your environment supports, as well as the model.
Control
Review access, permissions and approvals. A new capability does not authorise an action.
Articles attribute announcements to their providers. Use cases are evaluation proposals, not results from our own benchmarks or a list of models available in Hub.
2026 announcements
16 articles in EnglishEnterprise context in Gemini: when to use it and when to look beyond it
Google distinguishes operational work from idea exploration when choosing the sources AI should use.
Antigravity shows how skills can support a scientific workflow
Google DeepMind presents a chemistry workflow combining documents, data, tools and review.
OpenAI brings agent evaluation into physical-world engineering
The company demonstrates design, electronics and simulation work. Professional review remains part of the process.
FlashProve: proving which model replied does not prove its answer is correct
Anthropic publishes a verifiable-inference prototype. Its guarantee concerns computation, not factual accuracy.
Gemini 4 Carbon: new capabilities and a phased rollout
Google announces Carbon for engineering and cybersecurity, beginning with a limited group of partners.
GPT-6.1 Terra: understanding screens to work with applications
OpenAI presents a model focused on visual reasoning, computer use and writing. We suggest how a team could evaluate it.
Claude Haiku 5.5: what it offers for repetitive business tasks
Anthropic positions Haiku 5.5 for frequent, cost-sensitive work. The right choice depends on errors and review time for each task.
Mistral Large 4: a new preview for business evaluation
Mistral announced a public preview of Large 4. We explain that status and how to prepare a business-use comparison.
Gemini 4 Argon: complex work and expanded access
Google announced Argon in September and expanded access in October. Assess it through concrete tasks and verifiable results.
GPT-6.1 Sol: evaluating the cost of completed professional work
OpenAI launched GPT-6.1 Sol for coding and professional work. Assess its value through accepted deliverables and review effort.
Claude Sonnet 5.5: everyday documents and tasks under review
Anthropic positions Sonnet 5.5 for well-defined work. We propose evaluating it with documents grounded in approved information.
Claude Opus 5.5: evaluating complex tasks that require judgement
Anthropic announced Opus 5.5 for complex work. Clear goals, sources and review points help assess its contribution.
DeepSeek V4.1 Flash: visual understanding and efficient workflows
DeepSeek added native visual understanding to V4.1 Flash. Test its usefulness with images and data from everyday work.
GPT-6 Astra: what to evaluate in an end-to-end task
OpenAI launched Astra for demanding multi-stage work. Evaluate the final deliverable, sources and required review.
Gemini 3.5 Flash: a 2026 launch focused on agent workflows
Google introduced Gemini 3.5 Flash in May for agents and coding. Turn the announcement into a useful team evaluation.