Claude Code Gains a Built-in Browser, GPT-5.6 at OpenAI, Amp 50% Cheaper

Three eventful days for developer tooling and models. Anthropic is adding a built-in browser to Claude Code on desktop, OpenAI is rolling out the GPT-5.6 family (with a variant up to 25 times cheaper), and coding agent Amp is becoming about 50% cheaper by replacing Opus with GPT-5.6. Also included are Cursor’s Side Chats, HeyGen’s transparent WebM output, extended access to Claude Fable 5, and the general availability of GitHub Issue Fields.

Claude Code: a built-in browser on desktop

July 10 — Claude Code on desktop now includes a built-in browser (in-app browser). The agent can open documentation, mockups, or any website, then read, click, and interact with the page just as it already does with your local development servers. An important point regarding security: the environment is sandboxed and configurable—you decide whether browsing sessions persist.

Claude Code on desktop now has an in-app browser.

Claude can pull up docs, designs, or any other site. It can read, click through, and interact the same way it does with your local dev servers.

It’s sandboxed and configurable: you choose whether sessions persist. — @ClaudeDevs on X

The feature requires updating the desktop application to its latest version.

🔗 Release note — @ClaudeDevs

OpenAI Unveils GPT-5.6

July 10 — OpenAI introduced GPT-5.6, describing it as a major advance in “health intelligence.” The central argument is its performance-to-cost ratio: across the entire family, OpenAI claims better performance at a lower cost, with a GPT-5.6 Luna variant that outperforms GPT-5.5 (at its highest reasoning setting) while costing 25 times less.

GPT-5.6 is a major step forward for health intelligence.

Across the lineup, we’re delivering stronger performance at lower cost: GPT-5.6 Luna outperforms GPT-5.5 at its highest reasoning setting while costing 25x less. — @OpenAI on X

The family also includes a high-end variant, GPT-5.6 Sol Ultra. A post shared by the @OpenAI account states that this model, made available the previous day, reportedly produced a proof of the Cycle Double Cover Conjecture, a 50-year-old problem, using 64 subagents in just under an hour.

GPT-5.6 Variant	Announced Feature
GPT-5.6 Luna	Outperforms GPT-5.5 (maximum reasoning setting) at 25 times lower cost
GPT-5.6 Sol Ultra	General availability; reported proof of the Cycle Double Cover Conjecture (64 subagents, under one hour)

🔗 Mathematical proof shared by OpenAI — @__eknight__

Amp Switches to GPT-5.6 and Becomes 50% Cheaper

July 11 — After a period of heads-down development, coding agent Amp is back with a fundamental change: Quinn Slack (@sqs), CEO of Sourcegraph, announced that Amp is now about 50% cheaper on average after largely replacing Opus with GPT-5.6 under the hood.

Amp is now ~50% cheaper on average (and much better) since we mostly replaced Opus with GPT-5.6 and shipped this little guy. Wild times. — @sqs on X

Two other updates accompany this shift. The reasoning levels have been renamed low / medium / high / ultra (instead of smart / deep), a scale considered easier to understand. And “orbs” (the machines that run the agent remotely) can now make commits on your behalf rather than under Amp’s name, with a co-authored-by mention—useful, for example, when pushing to a repository where only recognized committers trigger a deployment.

🔗 Low/medium/high/ultra reasoning levels — @sqs 🔗 Orbs: commits in your name — @sqs

Cursor: Side Chats and Conversation Search

July 10 — Cursor released a set of updates focused on managing agent conversations. The main feature, “side chats,” lets you ask a question or explore an idea in parallel without interrupting the main conversation; each side chat is a persistent agent conversation that can be recalled in the main thread via an @-mention.

Cursor Update	Description
Side chats	Persistent agent conversations running alongside the main thread, recallable via an @-mention
Transcript search	Local search index for finding thousands of agent conversations
Project/repository selectors	Simplified and more powerful, allowing you to launch an agent in fewer clicks
Cloud agent hooks	New hooks for cloud agents

🔗 Cursor Changelog — Side Chats and Conversation Search

HeyGen: Transparent WebM Output for Avatars

July 11 — HeyGen is adding transparent WebM output to its API, CLI, and MCP server. By passing output_format: "webm" to the /v3/videos endpoint, the generated avatar is returned with a transparent background and a true alpha channel, without a green screen to remove.

The benefit is directly tied to integration: the team demonstrates an avatar embedded in a Minecraft world using three.js, where the transparent video is simply a video element overlaid on the canvas, alpha channel included. This makes it possible to insert an AI presenter into a web interface or rendering engine without a post-production step.

🔗 Transparent WebM output — @HeyGen

Claude Extends Access to Fable 5

July 12 — Anthropic is extending access to Claude Fable 5 across all paid plans and keeping Claude Code’s weekly limits 50% higher through July 19. As before, each user can devote up to half of their weekly limit to Fable 5; beyond that, they can continue using usage credits or switch to another model to remain within their limits.

🔗 Extended access to Fable 5 — @claudeai

GitHub Issue Fields Generally Available

July 12 — GitHub is making Issue Fields available to everyone. The feature lets users add structured, typed metadata to issues—priority, effort, dates, custom values—across multiple repositories and read this information at a glance without opening each issue.

🔗 GitHub Issue Fields generally available — @github

In Brief

GPT-Live rolled out worldwide — GPT-Live is now available to all ChatGPT users worldwide, with voice usage limits temporarily doubled for the weekend. 🔗 Announcement (shared by OpenAI)
GitHub Copilot code review: better instructions, not better tools — GitHub explains how simply rewriting the agent’s instructions (instead of changing tools) reduced the average review cost by about 20% while maintaining the same quality. 🔗 Article
CodeQL 2.26.0 — added support for Kotlin 2.4.0 and AI-powered prompt injection detection. 🔗 Changelog
GitHub Multilingual Repositories Dataset — an open-source dataset covering more than 40 million repositories and 80 million classification rows, mapping non-English READMEs, issues, and PRs (Korean dominates issue text, while Portuguese leads READMEs with more than 3 million repositories). 🔗 Announcement — @github
Gemini: study notebooks — Google explains how to create study notebooks in the Gemini app. 🔗 Google Blog
OpenAI Bio Bug Bounty — the biological vulnerability hunting program is becoming a permanent private program, with rewards doubled. 🔗 Announcement — @OpenAI

What It Means

The agent environment is becoming a complete workstation. Claude Code’s built-in browser, Cursor’s Side Chats and transcript search, and even GitHub’s Issue Fields all point to the same trajectory: coding agents no longer merely edit files; they browse, remember, and structure context. Giving an agent a sandboxed browser opens up documentation and mockups without leaving the tool; indexing conversations treats agent history as a reusable resource.

Inference cost has become a top-tier product lever. GPT-5.6 emphasizes “more performance for less” (up to 25 times less for the Luna variant), and Amp is immediately acting on that: about 50% savings by replacing Opus with GPT-5.6. For agent vendors, the price per token is no longer merely a constraint but an architectural parameter optimized from one version to the next.

The application layer is decoupling from model providers. The fact that an agent like Amp can replace one provider’s model with another’s almost transparently for the user confirms that value is shifting toward orchestration (model selection, reasoning levels, remote execution). Meanwhile, Anthropic is extending access to Fable 5 and raising its weekly limits: a retention battle fought as much through generous quotas as through model quality.

Media generation is being industrialized for production pipelines. HeyGen’s transparent WebM output, with a true alpha channel that can be used directly in a web engine, illustrates the shift from demo to pipeline: AI video is now conceived as a component to integrate (compositing, embedding) rather than as a fixed final deliverable.