OpenAI introduces Frontier as Anthropic releases Claude Opus 4.6

OpenAI introduced Frontier on February 5, 2026. The platform is designed to help companies build, deploy, and manage AI agents across internal systems.¹ It provides shared context for agents and supports onboarding, feedback, permissions, and governance.¹ Frontier was available only to selected customers at launch. OpenAI said broader availability would follow over the next few months.¹

Frontier complements ChatGPT Enterprise and Business and existing developer tools rather than replacing them.²³ The platform is also designed to work with agents from other providers.⁴ That could help companies manage agents from different sources in one place. Whether Frontier will become a generally available packaged product remained unclear at launch.³

Anthropic introduced Claude Opus 4.6 on February 5, 2026.¹ The model targets coding, large codebases, and long-running tasks involving AI agents.¹ It is available through claude.ai, the Claude API, and major cloud platforms.¹ For the first time in the Opus line, it offers a context window of up to 1 million tokens, initially in beta.¹

Anthropic reports improvements in planning, code review, debugging, and searching large document collections.¹ That positions the model for teams working across extensive repositories and multistep workflows.¹ A larger context window, however, does not by itself establish reliable performance on every long input. Companies should test the model on their own tasks and check whether the additional context improves results.

Cursor introduced a research preview of its Long-running Agents on February 12, 2026. They are designed to plan larger coding tasks before working on them autonomously for longer periods. Cursor says they produce larger pull requests that require fewer follow-ups. Access is through cursor.com/agents and is limited to selected subscriptions.¹

The preview builds on Cursor’s research into “self-driving codebases,” in which multiple agents coordinate work on a software project.² Cursor reported a peak of more than 1,000 commits per hour across hundreds of agents during an internal week-long run.³ That figure indicates the scale at which the company is testing coordination. It is not an independent measure of code quality or of the performance users should expect from the preview.

Augment Code launched Context Engine MCP on February 6, 2026. It makes the company’s semantic code search available through the Model Context Protocol. MCP-compatible coding agents, including Claude Code, Cursor, and Codex, can use it to retrieve repository context. Augment offers a local mode through the Auggie CLI and a hosted MCP endpoint.¹²

The additional context is intended to help agents account for existing code when making changes. In its own evaluation, Augment reports an overall performance improvement of more than 70%, along with fewer tool calls and conversation turns.¹² These are not independently verified results and do not establish a general gain for every project. For companies, the practical appeal is adding repository search to agents they already use rather than switching coding tools.¹²

Perplexity introduced Model Council on February 5, 2026. The feature sends a query to three AI models at once and combines their responses into one answer. It shows where the models agree and where their responses differ.¹ At launch, Model Council is available on the web to Perplexity Max subscribers.¹

This goes beyond switching between individual models: users get a combined answer without submitting the same question repeatedly.¹ A separate model reviews the results and attempts to resolve conflicts.¹ On February 12, Perplexity added Memory support, allowing the participating models to draw on relevant personal context when needed.² The available announcements do not provide benchmark evidence that combined answers are more reliable than responses from individual models.¹²

A team with researchers from Meta, Imperial College London, and the University of Cambridge posted the preprint “Scaling Small Agents Through Strategy Auctions” on February 4.¹ It proposes Strategy Auctions for Workload Efficiency (SALE), in which small AI agents bid with short plans. A mechanism selects a plan based on cost and expected value, while shared auction memory is meant to improve later decisions.¹ The paper does not describe a public deployment.¹

The method addresses a limitation of conventional model routing: a task description may not predict whether an agent will succeed over a longer workflow. SALE evaluates proposed strategies rather than routing solely by task and model.¹ In the reported tests, total inference costs fell 35%, while the success rate edged above that of the strongest single agent.¹² Whether those savings carry over to production workflows remains untested.