Copilot Adds Carbon, Fable 5.2: Pro+ $39, Max $100 [2026]

GitHub added Anthropic's Claude Fable 5.2 and Google's Gemini 4 Carbon to GitHub Copilot in changelog entries dated November 9 and 10, 2026, making both models available to Copilot Pro+, Max, Business and Enterprise users in VS Code, Visual Studio and Copilot CLI and selectable for the Copilot coding agent. GitHub lists Pro+ at $39 a month and Max at $100 a month, with agent work now drawing down a pool of AI credits where one credit equals one cent, and both models bill at the provider's list price under GitHub's usage-based system. The additions land on a platform GitHub reorganized around agents, including a connected Copilot cloud experience announced October 28, 2026, and the general availability of Agentic Workflows, which compiles plain Markdown into GitHub Actions YAML.

GitHub put two of the newest frontier models into Copilot this week, and both are built to run as agents instead of finishing your next line. Gemini 4 Carbon arrived in GitHub Copilot https://buildapps.co.uk/signals/gemini-4-carbon-github-copilot-default-model-enablement/ next to Claude Fable 5.2, with GitHub’s changelog entries dated November 9 and 10, 2026. Both models are open to Copilot Pro+, Max, Business and Enterprise users in VS Code, Visual Studio and Copilot CLI, and both can be picked for the Copilot coding agent. The price of entry is clear. GitHub lists Pro+ at $39 a month and Max at $100 a month, and agent work now draws down a pool of AI credits where one credit equals one cent. That shift in billing matters as much as the model names, because a long-running agent can burn through an allowance in a way a chat window never did. For a side-by-side of what each tool costs under load, see this Claude Code, Cursor and Copilot pricing comparison https://shattered.io/claude-code-vs-cursor-vs-github-copilot-pricing-limits-2026/ . Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. Add now https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F GitHub Copilot Agentic Coding Moves Past Autocomplete For most of its life, Copilot meant ghost text in an editor and a chat pane beside it. GitHub’s own materials now treat that description as outdated. On October 28, 2026, the company announced a connected Copilot cloud experience, and the rollout began the same day. Under it, a developer can start a cloud agent session from a phone, steer it from VS Code, and check the result on GitHub.com. That is a different product shape. The agent no longer lives inside one editor window. Sessions persist, they run in the cloud, and several clients attach to them. Free users keep Copilot Chat on GitHub.com and GitHub Mobile, but the heavy agent features sit behind paid plans and a credit meter. The same day, GitHub made Agentic Workflows generally available. The open-source GitHub Next project lets a team describe automation in plain Markdown, which compiles into GitHub Actions YAML. Tasks such as issue triage, CI-failure analysis, documentation updates and test writing run inside Actions, using Copilot or another coding agent. It is the clearest sign that GitHub wants agents running on its infrastructure, not just in a developer’s editor. So the model additions of the past week land on a platform that was already reorganizing around agents. GitHub Copilot agentic coding now spans GitHub itself, VS Code, Visual Studio, Copilot CLI, the Copilot app and Actions, with JetBrains and Xcode integrations reported alongside. Anyone weighing the product against Anthropic’s terminal agent should read this head-to-head on GitHub Copilot versus Claude Code https://shattered.io/github-copilot-vs-claude-code-2026/ , since both now offer Fable 5.2 to paying users. What GitHub Shipped on November 9 and 10 Two changelog entries carry the news. The first marks Claude Fable 5.2 generally available in Copilot for complex, multistep coding across large codebases. The second does the same for Gemini 4 Carbon. Both entries name the same four plans Pro+, Max, Business and Enterprise and the same surfaces. GitHub also says both models bill at the provider’s list price under its usage-based system. Anthropic released Fable 5.2 on November 9 together with Claude Mythos 5.2. The company calls Fable 5.2 its most capable model for demanding technical work. Our earlier report on Fable 5.2 and Mythos 5.2 https://shattered.io/anthropic-fable-5-2-mythos-5-2-30-percent-faster-8-40-pricing-2026/ covers the launch itself, so the focus here is what the Copilot rollout changes. Google’s side of the timeline is messier. It first announced Carbon on October 28 and limited early access to a group of cybersecurity defenders through its Fairwind program. On November 10, it moved to broad availability through the Gemini API in Google AI Studio and Android Studio, with enterprises reaching it through Gemini for business. Google AI Ultra subscribers got it in AI Studio and Google Antigravity, and a rollout to the Gemini app and AI Mode in Search began the same day. Two more items round out the week. Google announced Gemini 4 Hydrogen on November 12 as a faster-execution model for agentic workflows, and it framed Hydrogen as something other than a replacement for Carbon. GitHub’s model documentation also shows how crowded the picker has become. It lists Claude Opus 4.8 in fast mode as a preview, Gemini 4 Argon at promotional pricing through December 1, and several GPT-6 and GPT-6.1 variants Astra, Sol, Terra and Luna . One retirement is worth flagging for anyone with older habits. GitHub retired Claude Sonnet 4.6 on September 1, 2026. It remains available to individual subscribers on annual Pro and Pro+ plans, but not to monthly subscribers. Claude Fable 5.2 in Copilot: Price, Context and the Data Policy Anthropic prices Fable 5.2 at $8 per million input tokens and $40 per million output tokens. That is a cut from the $10 and $50 charged for Fable 5.1, and VentureBeat framed the change https://venturebeat.com/technology/anthropic-releases-claude-fable-5-2-and-mythos-5-2-pairing-new-record-high-benchmarks-with-20-api-cost-reduction as a 20% API cost reduction. The model’s documentation on the Claude Platform https://platform.claude.com/docs/en/models/fable-5-2/overview lists a 1-million-token context window and a 128,000-token maximum output. Cache reads cost $0.20 per million tokens, and batch processing carries a 50% discount. Anthropic’s own coding numbers come with the usual caveat that they are vendor figures. The company reports 49.1% on its FrontierCode 2.0 evaluation at Xhigh effort, against 32.4% for Fable 5.1. At High effort, it reports 45.3% for Fable 5.2, compared with 45.5% for Opus 5.5 at Max effort and 42.8% for GPT-6 Astra at its best reported score. Read that spread carefully. Fable 5.2 at High effort roughly matches Opus 5.5 at Max, which is the real argument for the newer model: similar results at a lower effort setting. Anthropic also says the model can work on ambiguous problems for days, check its own output, and orchestrate subagents in a loop. It describes Fable 5.2 as its strongest lead agent and says it can coordinate hundreds, even more than 1,000, subagents. Anthropic adds that more than 90% of the changes merged into its own core codebase are now written by Claude with human guidance. None of that is independent confirmation of how the model behaves inside Copilot, so teams should test on their own repositories first. The part Copilot administrators need to read is the data policy. GitHub’s changelog says Fable 5.2 requires data retention by default so that Anthropic’s safety classifiers can operate. Zero data retention is available only for certain eligible enterprise customers. On Business and Enterprise, an administrator must switch on the Claude Fable 5.2 policy in Copilot settings, and it is off by default. Leaving it off keeps the model unavailable to the whole organization. GitHub’s documentation also lists Fable 5, 5.1 and 5.2 among the models not covered by its own data retention agreement. Safeguards add one more wrinkle. In most Claude applications, queries flagged by the cybersecurity safeguards route to Opus 4.8, and biology flags route to Opus 5. API customers have to configure a new Fallback API. Anthropic reported about 55% fewer blocks with Fable 5.2 than with Fable 5.1 in early testing, with roughly 60% fewer in cybersecurity and 40% fewer in biology. Gemini 4 Carbon in Copilot: Cheaper Now, Pricier From December Carbon’s headline is price. Google lists an introductory rate of $2 per million input tokens and $10 per million output tokens through November 30, 2026. Cached input tokens get a 95% discount. From December 1, the announced standard price becomes $4 for input and $20 for output. Any team that budgets against the introductory number is building on a figure Google has already said will double. Google’s developer materials describe a 1-million-token context window and adjustable low, medium and high thinking levels. On the security side, Google says Carbon reaches 88.2% on SEC-bench Pro, which it calls state of the art for autonomously finding real flaws in real-world software. That fits the model’s origin as a tool first offered to defenders, and it makes Carbon a natural candidate for the security-review agents teams are wiring into pull requests. For a wider look at how it stacks up outside Copilot, see this breakdown of Claude Opus 5.5 versus Gemini 4 Carbon and Argon https://shattered.io/claude-vs-gemini-opus-5-5-vs-gemini-4-carbon-argon-59-vs-58-2026/ . A practical caution applies to the VS Code picker. One line in GitHub’s supported-models documentation marked Carbon’s availability in Visual Studio Code as “TBD” even as the changelog listed the editor. Documentation pages and changelogs drift apart in the first days after a launch. Open the model picker and check before you plan a rollout around it. Carbon’s billing in Copilot follows the same rule as Fable’s, meaning provider list price. GitHub confirms that principle, but the sources reviewed here do not show a separate Copilot rate for Carbon. Treat Google’s published API price as the reference until GitHub’s pricing table says otherwise. Fable 5.2 vs Gemini 4 Carbon: Spec Comparison | Item | Claude Fable 5.2 | Gemini 4 Carbon | |---|---|---| | Broad availability | November 9, 2026 | November 10, 2026 limited access from Oct. 28 | | Input price per 1M tokens | $8 | $2 through Nov. 30, then $4 | | Output price per 1M tokens | $40 | $10 through Nov. 30, then $20 | | Context window | 1 million tokens | 1 million tokens | | Max output | 128,000 tokens | Not stated in sources | | Reasoning control | Effort levels Medium default in Claude apps | Low, medium and high thinking | | Vendor-reported coding score | 49.1% FrontierCode 2.0 Xhigh | 88.2% SEC-bench Pro | | Copilot plans | Pro+, Max, Business, Enterprise | Pro+, Max, Business, Enterprise | | Org-level gate | Admin policy, off by default | Not stated in sources | The two benchmark rows do not measure the same thing, so do not read them as a head-to-head result. FrontierCode 2.0 tests agentic coding, while SEC-bench Pro tests vulnerability discovery. Giallo, a studio that compared the two models on November 9, read Anthropic’s own comparison table as showing Fable 5.2 ahead of Carbon on six of seven comparable tests, with Carbon winning AutomationBench at 53.9% against 51.4%. That is a reading of vendor data, not an independent test. The price gap is large and temporary. At the introductory rate, Carbon costs a quarter of Fable 5.2 on input and a quarter on output. After December 1, the gap narrows to half. Choosing between them will come down to how many agent turns a task needs, because a cheaper model that needs twice the retries gives the savings back. Our report on the Claude, ChatGPT and Gemini price gap https://shattered.io/claude-vs-chatgpt-vs-gemini-price-gap-2026/ shows how quickly these differences move. VS Code Becomes the Control Room for Agents The editor side of the story is where daily use changes. At GitHub Universe, VS Code announced Agent Host Protocol 1.0 on October 28. It lets agent sessions run independently of the editor and be reached from VS Code, GitHub Copilot, GitHub Mobile, github.com and Copilot CLI. VS Code also made its new Copilot harness the default for new sessions in the editor and the Agents window. The harness combines the local VS Code and Copilot CLI harnesses and sits on the Microsoft Copilot SDK developed by GitHub. Multiple Harnesses, One Window VS Code reported support for Claude and Codex harnesses through the same Agent Host, and each provider keeps its own runtime and capabilities. Users can sign in with a supported account or bring an API key. Copilot Fusion can pick one model or orchestrate several for a task, and a Compare Models view was available in preview in the Agents window. Next Edit Suggestions are on by default, and a modernized UI, previewed on October 15, became the default with an option to return to the classic one during the transition. Cloud sessions can continue after you close a laptop. Remote agents keep working on a machine, but only while that machine and its host stay running. That distinction matters for anyone planning to leave a long job overnight on a desktop that might sleep. What VS Code 1.145 Reportedly Adds According to GitHub’s weekly release notes for the November 2 cycle, published November 6, Copilot in VS Code 1.145 adds experimental chat views that collapse or expand reasoning and tool calls. It also adds previews of remote applications and interactive canvases, remote-browser support, and local automations scheduled with cron expressions. Files and screenshots can be attached through Add Context, and the release improves rendering of earlier terminal output. The weekly notes also mention workspace custom agents. VS Code’s updates page already carries a 1.146 heading, so check which build you are on before relying on a specific 1.145 feature. Cloud Copilot Automations, announced October 28, move scheduled work out of the laptop. Cron-based local automations and cloud automations give teams two places to run the same kind of recurring job, such as a nightly dependency check or a morning triage pass. Visual Studio, JetBrains and Xcode Get Their Share Visual Studio users get the same two models, and Microsoft’s agent features there build on work from earlier in the year. On February 17, 2026, the Copilot coding agent became delegable from Visual Studio. On April 2, Microsoft announced custom agents defined in .agent.md files, reusable agent skills, MCP connections, a find symbol tool and a Profiling Agent that points out performance bottlenecks and guides fixes. Microsoft’s documentation describes Visual Studio’s agent mode as taking a high-level natural-language task, working out steps, editing code, running terminal commands, invoking tools and applying changes across a codebase. The product is now simply Visual Studio 2026, and Copilot is described as integrated with awareness of the solution, repository and tools. Pricing for the IDE itself sits on a separate page with subscription tiers, and the Dev Essentials program remains a free route to Visual Studio Community and VS Code. Beyond Microsoft’s own editors, GitHub’s changelog says developers can now work with Copilot directly in Xcode’s built-in coding assistant. The Xcode integration is in public preview and starts with Xcode 27.3 beta. GitHub’s release notes also point to Copilot updates involving JetBrains Air, plus changes for pull requests, settings and agent sessions, though the details come from a single weekly post and deserve a second look in your own IDE. Review is the other front. Teams already running Copilot on pull requests should look at this walkthrough on setting up Copilot code review https://shattered.io/code-review-tools-copilot-setup-12-steps-80-min-2026/ , because a model that reads code as carefully as Carbon claims to is most useful on the review step, where a missed flaw costs the most. AI Credits: How Copilot Now Bills Agent Work GitHub’s billing page explains the unit plainly. GitHub AI Credits pay for AI usage in Copilot, every plan includes a monthly allowance, and one credit equals $0.01. When usage passes the included allowance, extra usage bills in credits at per-token rates shown in GitHub’s pricing tables. A user can set a dollar budget for additional usage, and a $10 budget covers 1,000 credits. Alerts fire at 75%, 90% and 100% of any configured budget. Teams that want a hard stop can create a bundled AI credits budget, set it to $0, and enable the option to stop usage when the limit is reached. Business and Enterprise allowances are pooled at the billing-entity level, while individual plans carry their own. Billing for cloud sandbox usage began in November 2026, and Free usage draws from a shared compute entitlement. Reports also describe agent mode, code review, Copilot in the cloud, Copilot CLI and the Copilot app as credit-metered, though that list comes from a single pricing report. | Plan | Monthly credits listed | Fable 5.2 and Carbon access | Notes | |---|---|---|---| | Free | Shared compute entitlement | Not among the plans GitHub names | Keeps Copilot Chat on GitHub.com and GitHub Mobile | | Pro | $15 | Not among the plans GitHub names | Positioned for everyday coding with agents | | Pro+ $39/month | $70 | Yes | Positioned for complex work with premium models | | Max $100/month | $200 | Yes | Highest individual allowance | | Business | Pooled per billing entity | Yes, after admin enables policy | Fable 5.2 policy off by default | | Enterprise | Pooled per billing entity | Yes, after admin enables policy | Zero data retention only for eligible customers | GitHub attaches a footnote to the monthly credit totals on its plans page, so read the fine print before you compare allowance to sticker price. The practical point stands regardless. A Pro+ subscriber who points Fable 5.2 at a large refactor can exhaust a $70 allowance faster than a month of chat ever would, and the budget controls are the only brake. What Developers Say: The Stack Overflow Data Adoption data helps size the audience for these features. The numbers below come from Stack Overflow’s 2025 Developer Survey, so they predate the November launches, and they describe developers in general, not Copilot customers. On agent adoption, the survey’s press release states that “31% of respondents currently use AI agents, while 17% plan to use them,” according to Stack Overflow’s 2025 survey summary https://stackoverflow.co/company/press/archive/stack-overflow-2025-developer-survey . The same release reports a favorable productivity signal from those users: “Among developers who have used AI agents at work, 69% agree that the agents increased their productivity.” Stack Overflow is the speaker in both cases, as cited research. The editor split matters just as much for Microsoft’s strategy. The 2025 Developer Survey https://survey.stackoverflow.co/2025/ reports that “Visual Studio Code is used by 75.9% of respondents, while Visual Studio is used by 29%.” Copilot’s agent push therefore rests mostly on VS Code, with Visual Studio as the second front. Those figures likely explain why VS Code gets the experimental features first. Read together, the numbers describe a market where a minority already uses agents and a further slice plans to. A product that makes agents easier to start, easier to leave running and easier to check from a phone is chasing that 17% plus the many developers who have not yet tried one. It also faces the usual limit of survey data: the 69% productivity figure reflects self-reported opinion, not measured output. Competitive Comparison: Copilot vs Claude Code and Other Agents Copilot’s pitch is choice. It offers a model picker spanning Anthropic, Google and OpenAI, and VS Code can host Copilot, Claude and Codex harnesses side by side. Anthropic’s pitch is depth. Claude Code uses Fable 5.2 as its strongest lead agent, with Medium effort as the default in Claude apps and Claude Code, and a fast mode that Anthropic claims delivers up to 2.5 times faster token generation at twice the standard price. The distinction shows up in cost control. Copilot meters everything through AI credits that draw down at a cent each, with budget alerts and a hard stop available. A team that wants one invoice across several model vendors benefits from that. A team that wants the model maker’s own agent tuning might prefer going direct. Our look at OpenCode versus Claude Code https://shattered.io/opencode-vs-claude-code-10-go-vs-20-pro-2026/ shows the open-source end of the same trade. Market share tells a similar story from another angle. Developer usage data for Anthropic’s and OpenAI’s terminal agents is covered in our Codex versus Claude Code usage comparison https://shattered.io/codex-vs-claude-code-stack-overflow-65-5-vs-29-5-2026/ , and it shows how contested the category is. Copilot’s advantage is distribution, since it sits in the editor three in four Stack Overflow survey respondents already use. The main approaches split like this. | Approach | Model access | Where agents run | Cost control | |---|---|---|---| | GitHub Copilot | Multi-vendor picker Anthropic, Google, OpenAI | Local, cloud sessions, Actions | AI credits at $0.01, budgets, alerts at 75%, 90%, 100% | | Claude Code | Anthropic models, Fable 5.2 available | Terminal and Claude apps | Plan limits, optional fast mode at 2x standard price | | VS Code Agents window | Copilot, Claude and Codex harnesses | Local and remote machines, cloud sessions | Depends on the harness and account or API key | | Self-hosted assistants | Open-weight or bring-your-own models | Your own hardware | Hardware and electricity, no per-token meter | Historical Context: How Copilot Reached This Point in 2026 The agent push did not start this week. Over the past nine months, GitHub and Microsoft have moved Copilot step by step from a suggestion engine toward a system that plans, executes and reports back. Dated, confirmed announcements trace the path. | Date 2026 | Milestone | |---|---| | February 17 | Copilot coding agent made delegable from Visual Studio | | April 2 | Custom agents in .agent.md files, skills, MCP and Profiling Agent arrive in Visual Studio | | September 1 | Claude Sonnet 4.6 retired, except for annual Pro and Pro+ subscribers | | October 15 | Modernized VS Code UI becomes the default, classic option kept | | October 28 | Connected cloud experience, Agent Host Protocol 1.0, Agentic Workflows generally available, cloud Automations | | November 9 | Claude Fable 5.2 generally available in Copilot | | November 10 | Gemini 4 Carbon generally available in Copilot, Google opens developer access | | November 12 | Google announces Gemini 4 Hydrogen for agentic workflows | The model churn is the other half of the history. Anthropic shipped two frontier models in quick succession, a pace we examined in our report on Anthropic’s two frontier models and the slowdown call https://shattered.io/anthropic-two-frontier-models-since-amodei-slowdown-call-ipo-2026/ . Copilot’s picker now absorbs those releases within days, which is a change from the era when a new model took weeks to reach an editor. The billing history matters too. A flat monthly price once covered nearly everything Copilot did. The move to AI credits turns the product into a metered service for agent work, much like a cloud account. That mirrors what happened to cloud computing years ago, when teams learned to watch a dashboard instead of a price list. Market Impact: Winners, Losers and Open Questions The model vendors gain distribution. Anthropic and Google both get their newest model in front of Copilot’s user base within days of launch, billed at their list prices. For Google, that is a route into a developer audience that chooses among several vendors. For Anthropic, it is one more channel where Fable 5.2’s price cut can drive volume. Teams get leverage but also complexity. A multi-model picker lets an engineering lead route cheap tasks to a cheaper model and hard tasks to a stronger one. It also means every developer can pick a different model, which makes cost forecasting harder. The admin policy on Fable 5.2 is a sign of what is coming, since each new model may carry its own data-handling terms that an organization has to approve one by one. Independent tools face a squeeze. Editors and agents that cannot match Copilot’s reach or a model maker’s own tuning must compete on openness or price, which makes self-hosted AI coding assistants https://shattered.io/self-hosted-ai-coding-assistant-qwen-code-2026/ a logical option for teams worried about metered bills. Our coding and developer tools coverage https://shattered.io/category/software/ follows that side of the market as well. Open questions remain. GitHub has not published a separate Copilot rate for Carbon in the sources reviewed here. The VS Code availability note for Carbon looked inconsistent across pages. And no independent benchmark yet shows how either model performs inside Copilot’s agent loop, as opposed to vendor-run evaluations. Until those arrive, treat every score in this piece as a claim to test, not a result to rely on. Five Predictions for Copilot’s Agent Push Analysis, not reporting, so none of these carry dates or numbers beyond those already confirmed. - Carbon’s December price will test loyalty. Once the introductory rate ends on November 30, teams that picked Carbon for cost will rerun their math, and some will shift routine tasks to other models in the picker. - Per-model admin policies will multiply. Fable 5.2 already needs an opt-in on Business and Enterprise. Expect other frontier models to arrive with similar gates, and expect security teams to write standing rules for them. - Budget controls become a buying criterion. The $0 hard-stop budget and the 75%, 90% and 100% alerts will matter more to finance teams than any benchmark score. - VS Code stays the lead surface. With Agent Host Protocol and multiple harnesses, new agent features will keep landing in VS Code first, with Visual Studio, JetBrains and Xcode following. - Review agents grow faster than writing agents. Security-tuned models such as Carbon fit the pull-request step, and review is where most teams can measure a clear gain without handing over production code. The takeaway: GitHub Copilot agentic coding is now a multi-model, metered, cloud-backed service, and the new Fable 5.2 and Carbon options make cost and policy decisions as important as model quality. Frequently Asked Questions Which Copilot plans include Claude Fable 5.2 and Gemini 4 Carbon? GitHub lists both models for Copilot Pro+, Max, Business and Enterprise users. They appear in the model picker in VS Code, Visual Studio and Copilot CLI, and in the Copilot coding agent. Business and Enterprise administrators must enable the Fable 5.2 policy first. How much do Fable 5.2 and Carbon cost? Anthropic prices Fable 5.2 at $8 per million input tokens and $40 per million output tokens. Google’s introductory Carbon price is $2 and $10 through November 30, 2026, rising to $4 and $20 from December 1. GitHub says both bill at the provider’s list price. What is an AI credit in Copilot? One GitHub AI Credit equals $0.01. Each plan includes a monthly allowance, and extra usage bills in credits at per-token rates. A $10 budget covers 1,000 credits, and alerts fire at 75%, 90% and 100% of a configured budget. Does Fable 5.2 keep my code? GitHub says Fable 5.2 requires data retention by default so Anthropic’s safety classifiers can run. Zero data retention is available only to certain eligible enterprise customers. Read GitHub’s retention documentation before enabling the policy for your organization. Can I run agents on my own machine and in the cloud? Yes. Cloud sessions keep going after you close a laptop, while remote agents keep working only as long as their host machine stays running. VS Code’s Agent Host Protocol lets sessions run independently of the editor. Is Claude Sonnet 4.6 still available in Copilot? GitHub retired Claude Sonnet 4.6 on September 1, 2026. It remains available to individual subscribers on annual Copilot Pro and Pro+ plans, but not to monthly subscribers. What is Gemini 4 Hydrogen, and does it replace Carbon? Google announced Hydrogen on November 12, 2026 as a faster-execution model for agentic workflows. It is not a replacement for Carbon.