Weeks of local AI testing: local models hit only 20-35% of Claude 5.5 coding productivity
srchvrs · x · 2026-10-12
After weeks of testing local AI setups for full app builds, Julian Harris concludes:
- Intelligence is good enough: local models (on up to 128GB hardware) are close enough to frontier models to be an everyday substitute.
- Speed is the blocker: productivity tops out at 35% of the Claude 5.5 family, with most setups at 20-25%.
- Verbosity remains a problem for local models.
In follow-ups he cites throughput of 30-90 tokens/s across his hardware, notes concurrency and sub-agent handling differ across inference servers and harnesses, and estimates GPT 6 sol runs 70-80 tokens/s — still well ahead of local options.
Related event: Local AI Coding Is Smart Enough but Only 20-35% as Fast as Claude(2 posts)→
More from coding & agent
- Index Network open-sources Social Discovery Protocol: agents negotiate fit on your behalf — carsonfarmer · 2026-10-23
- Installed a formal-verification skill: it silently installed Java, MacBook runs 10 degrees hotter — zealcaiden · 2026-10-23
- OpenAI's Decisions API hits GA: near real-time routing up to 10x faster than GPT-6 Luna — OpenAIDevs · 2026-10-23
- Ruff author warns: upgrade Node/Bun now — --use-system-ca still accepts the eMudhra cert — charliermarsh · 2026-10-23
- Proof of Agent: paste one prompt and your AI mines, grades, and self-governs — adamamcbride · 2026-10-23
- DeepLearning.AI weekly: GPT-6.1 hits 97.9% byte parity on COBOL migration — DeepLearningAI · 2026-10-23