Post cover image

May 31, 2026

Stop Paying $2,000 a Month for Claude Code.

Microsoft just cancelled Claude Code for thousands of its own engineers because each developer was burning $500-$2,000 a month on tokens…

Abhishek Agarwal

7 min read

Stop Paying $2,000 a Month for Claude Code. The 24B Open-Source Model on a Single RTX 4090 Closes the SWE-bench Gap to 12 Points.

Microsoft just cancelled Claude Code for thousands of its own engineers because each developer was burning $500-$2,000 a month on tokens. This is the local-first stack that runs on a single RTX 4090, scores 68% on SWE-bench Verified, and costs you zero per token forever.

On May 15, 2026, Rajesh Jha sent an internal memo to thousands of Microsoft engineers across Windows, Microsoft 365, Outlook, Teams, and Surface: their Claude Code licenses were getting cancelled by June 30. The official reason was "benchmarking-then-convergence." The actual reason, surfaced by The Next Web and an internal Uber comparison, was simpler: engineers were spending $500 to $2,000 per month on API tokens, and the cost scaled with adoption.

Uber deployed Claude Code to roughly 5,000 engineers. By April 2026, monthly active usage hit 84–95%. The better the tool got, the more developers used it, and the higher the bill grew. That is the same paradox now sitting on every Fortune 500 CFO's desk.

The technical answer to this has existed quietly for six months and just hit a tipping point.

A $500-$1,500 GPU + a 24-billion-parameter open-source model now scores 68.0% on SWE-bench Verified. Claude Sonnet 4.6 scores 79.6% on the same benchmark — a 12-point gap. For most enterprise coding work, that gap is invisible. For your monthly bill, it is the difference between $24,000 a year per developer and zero.

Here is the stack.

What "good enough" actually looks like in 2026

Three open-source coding models cleared the bar in the last 90 days:

Devstral Small 2 (24B parameters, Mistral) — 68.0% SWE-bench Verified, 256K context window, Apache 2.0 license. Released as the consumer-friendly variant of Devstral 2, designed specifically for agentic coding. Runs on a single RTX 4090 (24GB VRAM) at Q4_K_M quantization, or any Mac with 32GB unified memory. Model card on Hugging Face.

MiniMax M2.5–80.2% SWE-bench Verified. Matches Claude Opus 4.6 (80.8%) within 0.6 percentage points. Open-weight. The gap between open-source and proprietary models is effectively closed for everyday coding work.

DeepSeek V4 Pro — Released April 24, 2026. Scores 69.99 on LiveBench Coding Average and 56.67 on Agentic Coding (May 12, 2026 snapshot). Strongest pure-coder economics in the open-source space, MIT-licensed weights.

If you want one model to start with: Devstral Small 2 is the safe pick. Apache-licensed, runs on consumer hardware, designed from day one for agentic loops with tool use and function calling baked in.

The hardware: $500–$1,500 one-time vs $24,000/year hosted

Three viable build paths for the developer who wants to escape Anthropic's token meter:

Path 1 — RTX 4090 desktop (~$1,500). 24GB VRAM. Runs Devstral Small 2 at Q4_K_M in roughly 14GB VRAM with 10GB headroom for KV cache (long context) and inference framework overhead. Expect 60–100 tokens/sec for 24B-class models. Used RTX 3090s at ~$700 work nearly as well if you can find one with good thermals.

Path 2 — RTX 5070 desktop (~$500-$700). 12GB VRAM. Tighter fit but workable with aggressive quantization (Q3_K_M) for 24B models, or comfortable with 9B-13B class models like Qwen 3.5 Coder 9B. Better suited if you want a budget build.

Path 3 — Mac with M4 Max / Mac Mini 48GB (~$1,500-$3,000). Unified memory makes 32B+ models trivially loadable. Mac MLX runs Qwen 3.5 Coder 32B at 60–70 tokens/sec versus ~35 tokens/sec through Ollama's llama.cpp backend, per InsiderLLM's benchmarks. The Mac Mini 48GB at roughly $1,500 loads a 32B model that would otherwise need a $700+ used RTX 3090 plus a separate workstation.

The economics, all in:

  • Hosted Claude Sonnet 4.6: $500-$2,000 per developer per month → $6,000-$24,000/year/dev
  • Local stack: $500-$1,500 hardware one-time, $0/month after

For a 50-engineer team at $500/dev/month, the breakeven on a $1,500 RTX 4090 is three weeks. At $2,000/dev/month, it is five days.

The agentic layer: OpenCode is the new Claude Code

The reason Claude Code dominated the conversation in 2025–2026 was not the model — it was the orchestration: a terminal UI that planned, executed, ran tests, and committed. That same pattern now exists open-source.

OpenCode — crossed 150,000 GitHub stars and ~6.5M monthly active developers by mid-2026. Written in Go using the Bubble Tea TUI library. The key features:

  • Supports 75+ LLM endpoints including Anthropic, OpenAI, AWS Bedrock, Azure OpenAI, OpenRouter, Grok, and any OpenAI-compatible local endpoint including Ollama
  • --plan agent (read-only, will not touch files) versus --build agent (writes code and commits)
  • Native MCP server support for tool plug-ins
  • SDK for embedding in scripts
  • Provider-neutral by design — no vendor lock-in

Point OpenCode at a local Ollama endpoint running Devstral Small 2 and you have Claude Code's workflow without Claude Code's bill. opencode.ai | GitHub.

Aider — the git-native pick. Reads your codebase, makes edits, commits every change automatically. Repomap feature is still the best in class for large monorepos. Best if you want commit-per-edit workflow.

Cline — the VS Code-native pick. 5M VS Code installs as of May 2026, making it the most-adopted open-source coding extension on the planet. Lives inside the editor instead of a TUI.

All three hit local Ollama with one config line. Pick whichever matches your editor habit.

The full build (45 minutes, weekend project)

Step-by-step, copy-paste ready:

1. Install Ollama

curl -fsSL https://ollama.com/install.sh | sh

2. Pull Devstral Small 2

ollama pull devstral:24b

This downloads roughly 14GB. The Q4_K_M quantization is the sweet spot — 95%+ quality retention at four-times size compression versus full precision.

3. Install OpenCode

npm install -g opencode-ai

4. Point OpenCode at local Ollama

Set the OpenAI-compatible endpoint:

export OPENAI_API_BASE=http://localhost:11434/v1
export OPENAI_API_KEY=ollama
opencode auth login ollama

5. Run

cd your-project
opencode --build --model devstral:24b

That is the whole setup. You now have a Claude Code-style terminal agent running fully local, against your codebase, with zero per-token cost.

For a Mac-native build on Apple Silicon, swap Ollama for MLX to get 2x the inference speed:

pip install mlx-lm
mlx_lm.server --model mistralai/Devstral-Small-2-24B-Instruct-2512

Expect roughly 60–70 tokens/sec on an M4 Max.

Where local actually loses (be honest)

Three places where the hosted model still wins, and you should know them before you go all-in:

1. Codebase-scale agentic tasks. Claude Code's Dynamic Workflows (released May 28) orchestrates 1,000 subagents in parallel with a JavaScript runtime that externalizes state to script variables instead of bloating the model's context. That is not currently achievable with the local stack — you do not have the rate-limit headroom or the meta-orchestration layer. For repo-wide migrations across hundreds of thousands of lines, stay hosted.

2. Frontier reasoning workloads. Opus 4.8 jumped to 96.7% on USAMO 2026 math (up from 69.3% in Opus 4.7). Local 24B-class models will not match that on hard reasoning chains. For straight code completion, refactoring, and routine agentic work, the gap is mostly invisible. For deep algorithmic problem-solving, it is not.

3. Long context with high recall. Devstral Small 2's 256K context is real, but recall accuracy degrades past roughly 100K tokens — common across all current local models. Opus 4.8's 1M context with strong recall remains the leader for full-monorepo reading work.

If your day is 80% routine refactoring + 20% hard reasoning, the local stack wins on cost without measurable productivity loss. If your day is 80% hard reasoning + 20% routine, stay hosted — or run a hybrid where local handles the routine and a small hosted plan handles the rest.

The economics in one table

The full cost comparison, per developer:

  • Claude Code (Sonnet 4.6): $500-$2,000/month → $6K-$24K/year. SWE-bench 79.6%. 5 minute setup.
  • GitHub Copilot Pro: $19-$39/month + usage → $228-$468+ /year. SWE-bench ~60%. 2 minute setup.
  • Cursor Composer 2.5: $20/month base + token usage. SWE-bench ~78%. 5 minute setup.
  • Local: RTX 4090 + Devstral Small 2 + OpenCode: $0/month after $1,500 hardware. SWE-bench 68%. 45 minute setup.
  • Local: Mac M4 Max + Devstral Small 2 + OpenCode (MLX): $0/month after $3,000 hardware. SWE-bench 68%. 45 minute setup.

Single-developer breakeven at $500/month Claude Code usage: three months.

At $2,000/month: twenty-four days.

For a 50-engineer team averaging $1,000/month/dev: the team breaks even in under two weeks.

What to do this weekend

If you are a solo developer or a small team currently paying for Claude Code or Cursor at scale:

  1. Buy or repurpose a single RTX 4090 (or use an M-series Mac with 32GB+ unified memory)
  2. Install Ollama
  3. Pull Devstral Small 2
  4. Install OpenCode
  5. Point OpenCode at the local Ollama endpoint
  6. Run your normal workflow for a week against the local stack
  7. Measure honestly: how often did you actually need to fall back to hosted?

For most developers in 2026, the answer is less than 20% of the time — and that 20% can stay on a smaller hosted plan rather than a team-wide enterprise contract.

The reason Microsoft just cancelled Claude Code internally is the same reason you should look at this stack. The open-source layer of the AI coding stack is good enough, and getting better at a faster cadence than the hosted price cuts. Devstral Small 2 hit 68% on SWE-bench in May 2026. The next 24B drop is expected in Q3.

The bubble that funds $965 billion valuations does not pop when models stop working. It pops when developers realize the open-source ones are good enough.

If you want the bigger AI bubble framing — including why Microsoft really cancelled Claude Code internally and what Anthropic's $965 billion valuation actually rests on — read the full breakdown on Substack here.

One More Thing — Subscribe to noobprogrammer16 on Substack

If this piece resonated, I write a weekly Substack on the messy economics behind the AI industry — IPO timing, the enterprise CFO conversations Anthropic doesn't want you to read, hidden costs nobody is pricing in, the bubble math, the unbundling timelines, and the trades worth watching.

The latest piece is the full backstory on why Microsoft cancelled Claude Code internally and what Anthropic's $965 billion valuation actually rests on. It pairs directly with this one.

👉 Subscribe at noobprogrammer16.substack.com — free, weekly, no spam.

And if you want the companion post to this article: Microsoft Cancelled Claude Code the Week Anthropic Hit $965 Billion →

References