Edward Boyce’s Post

The era of "all you can consume" AI for developers is officially ending. Woke up to the news yesterday that GitHub Copilot starting June 1, 2026... is moving to usage-based billing. While Claude Code, Cursor and other tools have also followed. It's a fundamental shift in how we build with agents. I posed about this last year that the subsidization of LLM costs was not going to last too long. Here we are now, the compute demands have become unsustainable. A single agentic loop can burn more tokens than a developer used in an entire month under the old flat-rate model. For copilot this is what it will look like from June: - "Unlimited" is replaced by credits: Your $10/mo plan now gives you exactly $10 in "GitHub AI Credits." (Personal observation, I consume $10 easily in a 6-8 hours of use with Sonnet on Copilot) - Token-based billing: You’re paying for every input, output, and cached token you consume. - Code reviews will take from that budget and will also consume github runner minutes. Double whammy there. Why does this matter? Because it forces a move toward what I call "Efficient Agency." The old model, a good agent was one that eventually found the answer, regardless of how many tokens it burned. The new eval benchmark for the future will be solving the problem with the absolute minimum number of tokens. However I dont think this is a bad thing. This shift will finally flush out the "wasteful" agents that just loop until they hit a context limit. It's going to reward engineering craftsmanship over "vibe coding" loops. P.S. At Optimal AI, we’ve been obsessing over this for a while. We use smart model routing and multi-model techniques to keep quality high while keeping costs drastically lower. This is how we can continue to provide unlimited-style value in a usage-based world. #GitHubCopilot #AIEfficiency #EngineeringLeadership #LLMOps #OptimalAI

GitHub Copilot went from: "We can't take new users." To: "Pay per use." That's not a pricing update. That's a signal. When a product is capacity-constrained, it means demand outran infrastructure. That's a good problem. But it also means the old pricing model, flat subscription, unlimited usage, stopped making sense. Because some users were using a little. And some were using everything. Usage-based billing fixes that. The heavy users pay more. The light users pay less. The economics align with the value actually being delivered. But here's the more interesting implication. When AI coding tools move to usage-based pricing, the conversation inside every engineering org shifts. It's no longer "do we have Copilot?" It's "how much are we actually using it — and is the output worth what we're paying?" That's a harder question. And a healthier one. The teams that use it constantly and ship faster will justify the cost easily. The teams that had it running in the background, barely touched, on a flat subscription? Now they have to reckon with whether AI actually changed how they work. Or just felt like it did. Usage-based pricing doesn't just change what you pay. It forces honesty about what you got. #GitHub #Copilot #AI #Engineering #FutureOfWork

🚀 GitHub Copilot’s Pricing Shift: What the 9x Increase for Claude Models Means for Developers in 2026 AI communities are reacting to GitHub’s upcoming changes to Copilot billing: starting June, model multipliers for Claude (and others) are rising sharply under the new usage-based structure. What many are calling a 900% effective hike for Claude-powered workflows is prompting teams and individual developers to reassess their tooling stacks, especially those relying on frontier models for agentic coding and large-scale projects. Key themes from the discussion: End of the subsidized era — Flat-rate access that enabled heavy experimentation is giving way to more realistic compute economics as demand grows Practical alternatives — Strong early praise for Anthropic’s official Claude VS Code extension, which many report supports full-project context, change review/accept workflows, and smoother integration Strategic responses — Surge in interest for multi-model setups (Claude + Gemini + open-source), direct Anthropic subscriptions (Pro/Max tiers), and better cost monitoring to protect unit economics Broader implications — Heightened focus on visibility (token counters, spend estimates) and predictable pricing as agentic tools become core to development velocity This moment reflects a maturing AI tooling market in 2026. As capabilities advance, providers are aligning pricing with real infrastructure costs — pushing organizations toward disciplined evaluation, hybrid architectures, and workflows that maximize ROI rather than raw consumption. For engineering leaders, developers, and AI practitioners, the takeaway is clear: treat model access as a strategic resource. Successful teams will combine the right models with strong prompting practices, skills frameworks (like GSD or Superpowers), and observability to sustain productivity gains without budget surprises. Are you feeling the impact of these Copilot changes or already exploring alternatives like the official Claude extension, Cursor, or direct API integrations? What strategies — multi-model routing, usage dashboards, or open-weight fallbacks — are you implementing to keep costs manageable while maintaining momentum? Share your experiences or evaluation tips; practical insights like these help the community navigate this evolving landscape more effectively. #ArtificialIntelligence #GenerativeAI #Claude #GitHubCopilot #AICoding #DeveloperTools #AgenticAI #AICostManagement #VSCode #LLM

𝗦𝗮𝗺𝗲 𝗮𝗴𝗲𝗻𝘁𝗶𝗰 𝗰𝗮𝗽𝗮𝗯𝗶𝗹𝗶𝘁𝘆. 𝗧𝘄𝗼 𝘃𝗲𝗿𝘆 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲𝘀. 𝗩𝗲𝗿𝘆 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝗼𝘂𝘁𝗽𝘂𝘁𝘀. Here's what most people miss 👇 Both GitHub Copilot and Claude Code can plan, execute, and iterate across your full codebase autonomously. Both are genuinely agentic in 2026. So why do outputs feel so different? It's not just the model. It's 𝘸𝘩𝘢𝘵 𝘵𝘩𝘦 𝘱𝘳𝘰𝘥𝘶𝘤𝘵 𝘪𝘴 𝘣𝘶𝘪𝘭𝘵 𝘢𝘳𝘰𝘶𝘯𝘥. 𝗚𝗶𝘁𝗛𝘂𝗯 𝗖𝗼𝗽𝗶𝗹𝗼𝘁 𝗶𝘀 𝗽𝗹𝗮𝘁𝗳𝗼𝗿𝗺-𝗳𝗶𝗿𝘀𝘁. It's deeply integrated into the GitHub ecosystem — issues, PRs, CI/CD pipelines, GitHub Actions. It supports multiple models (GPT, Claude, Gemini) and lets you assign entire issues to a coding agent that works in the background and opens a PR for review. The power is in the workflow integration, not any single model. 𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝗶𝘀 𝗮𝗴𝗲𝗻𝘁-𝗹𝗼𝗼𝗽-𝗳𝗶𝗿𝘀𝘁. The entire execution cycle — reading your codebase, planning, running tests, iterating on failures — is built specifically around how Claude reasons. No IDE required. Works in your terminal, browser, or desktop. You can even spawn parallel agents coordinating on the same task. The power is in the depth of the reasoning loop. Same underlying model capability. Very different product design philosophies. The real insight: 𝗔𝗜 𝗺𝗼𝗱𝗲𝗹 𝗰𝗮𝗽𝗮𝗯𝗶𝗹𝗶𝘁𝘆 = 𝗔𝗜 𝗽𝗿𝗼𝗱𝘂𝗰𝘁 𝗰𝗮𝗽𝗮𝗯𝗶𝗹𝗶𝘁𝘆. Don't just ask which model a tool uses. Ask what 𝘵𝘩𝘦 𝘱𝘳𝘰𝘥𝘶𝘤𝘵 𝘪𝘴 𝘣𝘶𝘪𝘭𝘵 𝘵𝘰 𝘥𝘰 𝘸𝘦𝘭𝘭 — and whether that matches your workflow. 💬 Which one fits your workflow better — or are you running both? Drop your take below. #AI #ClaudeCode #GitHubCopilot #AIEngineering #DeveloperTools #AgenticAI

The latest changes to GitHub Copilot don’t just feel like a product update — they feel like a reality check. Tightened usage limits, reduced model access, and paused new sign-ups all point to one thing: the current model for AI developer tools may not be sustainable at scale. Yes, agentic workflows are more demanding. Yes, infrastructure costs are real. But from a user perspective, this shifts the burden back to developers — forcing us to think about tokens, limits, and model multipliers while trying to stay productive. That’s a step in the wrong direction. One of the biggest promises of tools like Copilot was to reduce cognitive load, not introduce a new layer of resource management. What’s more concerning is the signal this sends: If even a flagship product like Copilot needs to pull back on availability and tighten limits, what does that mean for the long-term viability of AI-assisted development as we know it today? Transparency improvements (like usage visibility in VS Code) are welcome — but they don’t address the core issue: 👉 The gap between how these tools are marketed vs how they actually scale under real-world usage. This feels less like a temporary adjustment, and more like the beginning of a pricing and capability correction across the industry. https://lnkd.in/gZuuVBE8 #GitHubCopilot

The most underrated AI dev tool right now is boring on purpose. GitHub Models Evals lets you turn prompts into a versioned artifact and run evals from the GitHub CLI against real test cases. It feels like the missing step between “works in the playground” and “safe to ship in prod”. What’s genuinely good is the workflow. Prompts live in the repo as a .prompt.yml, evals can run locally or in CI, and you get repeatable signals instead of vibes. Honest take. It will not design your evals for you, and it will not magically pick the right metrics. You still need to write cases that reflect your product’s failure modes. If you are shipping an LLM feature, start by codifying 10 real inputs, then wire one eval into CI before you touch prompt iteration again. At Mobi-soft, we like tools that reduce argument time and increase shipping time. If you want help choosing and shipping with the right AI tools, DM us. Follow the Mobi-soft page for fresh vibe coding tools and AI Product Development news. #AIDevTools #LLMOps #ProductEngineering #PromptEngineering #Evals #GitHub #AIProductDevelopment

The premise needs a small correction first: GitHub Copilot hasn't actually lost the AI coding war — but it has gone from dominant pioneer to a genuinely threatened incumbent. Here's the full picture: Copilot is struggling, but still standing Copilot holds 42% market share and has reached 20 million cumulative users, deployed across 90% of Fortune 100 companies. Quantumrun Those are not the numbers of a defeated product. But the competitive pressure is very real, and the cracks are showing. Why developers are losing faith: Developer complaints about suggestion quality, latency, and context awareness have grown significantly since late 2025, with model swaps being a primary cause — GitHub cycled through Codex, multiple GPT-4 variants, and GPT-5 series, and each transition introduced regressions. Nxcode In March 2026, Copilot injected promotional "tips" into over 1.5 million pull requests, badly eroding developer trust. Nxcode Copilot's suggestion acceptance rate sits at 35–40%, compared to Cursor's 42–45%. Nxcode Tasks requiring changes across 10+ files with architectural implications produce noticeably more mistakes than competing tools. Nxcode Its characteristic failure mode is the confident wrong answer.

Friday reminded me why I’m genuinely excited about where AI tooling is heading. I pointed GitHub Copilot to a separate repo — one that already had auth/API access code built out — and asked it to pull what I needed into my current project and wire up the matching endpoint. It found the repo. Found the code. Dropped it in. Built the scaffolding around it. What would have taken a developer a solid chunk of their day? Done in minutes. But here’s the part that doesn’t get said enough: 🧭 AI is a GPS, not a driver. A GPS gets you there faster. It reroutes when traffic hits. It saves you from guessing at every turn. But if the road is icy, if the bridge is out, if something just feels off — you still need someone with hands on the wheel and the experience to know what to do next. When that generated code doesn’t behave the way it should, when the integration breaks in a way the tool didn’t anticipate, when the edge case shows up at 4pm before a release — that’s not an AI problem to solve. That’s a developer problem. The win isn’t AI replacing the craft. The win is AI eliminating the crawl so skilled developers can spend their time on the parts that actually require them. Know your stuff. Use the tools. Get there faster. #SoftwareDevelopment #GitHubCopilot #AITools #DeveloperProductivity #TechLeadership

GitHub paused new Copilot signups this week. The reason: "a handful of agentic requests can now exceed the monthly plan price." That's not a compute problem. That's a subscription model designed for autocomplete trying to hold against agent workloads. Copilot was priced around fast, stateless suggestions. Agentic sessions are long-running, parallelized, and can burn through a month's budget in an afternoon. Response: freeze signups, gate Opus to premium tiers, add weekly token limits. All the same fix. Slow demand until pricing catches up. Usage-based makes economic sense but unpredictable bills kill adoption. Flat-rate is simple but bleeds at scale. The other major AI coding tools will face the same math eventually. https://lnkd.in/gkMkK-NF

In 2026, tokens are the new developer currency. Stop spending yours on "Orientation Tax." I built a VS Code extension that cuts GitHub Copilot's token usage by 40–95%. It's now live on the Marketplace. Here's the problem: Every time Copilot needs to understand your code, it reads entire files — thousands of tokens just to figure out where a function lives. That's like reading an entire book to find one chapter title. 𝗧𝗼𝗸𝗲𝗻𝗦𝗹𝗮𝘆𝗲𝗿 fixes this by giving Copilot compact structural skeletons instead of raw files: → Before: 1,200 lines of raw code = 5,000 tokens → After: 8-line structural skeleton = 200 tokens 𝗛𝗼𝘄 𝗶𝘁 𝘄𝗼𝗿𝗸𝘀: ⚡ Single LSP call extracts the full symbol tree 🧠 Language-specific compactors strip bodies, keep signatures 📦 Content-hash LRU cache — instant on repeat access 🔧 Registers as a Language Model Tool that Copilot calls autonomously 𝗧𝗵𝗲 𝗱𝗮𝘀𝗵𝗯𝗼𝗮𝗿𝗱 alone took a wild turn: 📊 Real-time donut chart + coverage ring 🏆 Top savers leaderboard with medals 📈 Session timeline sparkline 🛡️ Automatic secrets detection — blocks API keys, tokens, private keys from ever reaching the LLM 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲𝘀: TypeScript, JavaScript, Python, Go, Java, Rust 𝗧𝗵𝗲 𝗱𝗲𝘀𝗶𝗴𝗻 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻 𝘁𝗵𝗮𝘁 𝗺𝗮𝗱𝗲 𝗶𝘁 𝗳𝗮𝘀𝘁: I intentionally cut Call Graph Extractor, Type Hierarchy Extractor, and Query Matcher. This keeps overhead at exactly 1 API call per file instead of 60+. Sometimes the best feature is the one you don't build. 🔗 Install from VS Code Marketplace: https://lnkd.in/e8iY364q  📂 GitHub: https://lnkd.in/eZu_H9_B  📄 MIT Licensed | Built with TypeScript If you're spending money on AI coding tokens, this might save you a lot. #OpenSource #VSCode #GitHubCopilot #AI #DeveloperTools #TypeScript #LLM #TokenOptimization #CodingTools #SoftwareEngineering