FOR CTOS · UPDATED OCT 10, 2026
Ship on open weights. Own your stack.
Run open-weight models for the routine majority of engineering work (boilerplate, tests, schema changes, debugging) and keep a premium closed model for the hard architectural calls. Coding-focused open models like GLM-5.3 and Qwen3.8 now sit close to the closed frontier, and DeepSeek V4.1-Flash output costs 88% less than Claude Sonnet 5.5. Deploy privately wherever code or data is sensitive.
By Lucas Page, The ShanghAI Guy · Updated October 10, 2026
Free. Benchmarks, deployment playbooks and model picks every Monday.
TOKENS OF CONTEXT
1M
GLM-5.3, Qwen3.8-Max, DeepSeek V4 and Kimi K3 all ship a 1M-token context window, with open weights you can host yourself.
Sources: Zhipu, Alibaba, DeepSeek and Moonshot model cards, 2026.
THE ENGINEERING SHORTLIST
Which model for which engineering job
JOB
MODEL
$ IN / OUT PER 1M
WHY
Coding agents on hard tasks
GLM-5.3 (Zhipu)
$1.40 / $4.40
Top-tier open-weight coding model, open weights since August.
High-volume pipelines and CI jobs
DeepSeek V4.1-Flash
$0.30 / $1.20
Half price off-peak. Run nightly jobs there.
Cheapest capable assistant
Qwen3.8-Flash / 27B
$0.16 / $0.47
27B runs locally under Apache 2.0, Flash is the hosted tier.
Large repos, long agent loops
Kimi K3 (Moonshot)
$3.00 / $15.00
Strong on repo navigation, but pricier than Claude on output.
Managed enterprise code API
Seed 2.0 Code (ByteDance)
$0.50 / $3.00
Via BytePlus ModelArk, no infrastructure to run.
Reference
Claude Sonnet 5.5
$2.00 / $10.00
Keep it for the hardest refactors.
Official list prices, standard tier, checked October 10, 2026.
THE BENCHMARK
Same PR. 16x cheaper.
We ran an identical 12-file FastAPI refactor through Claude Code and OpenCode with Qwen. Claude needed zero corrections at $5.12. Qwen self-corrected one test fixture and finished for $0.31.
Tested in September 2026 on Qwen 2.5 Coder and earlier Claude pricing. Sonnet 5.5 now lists at $2 / $10, so the gap is smaller, but open models moved up too.
Engineering teardowns
![[W39] 5 new open-weight models to replace closed APIs this week](https://media.beehiiv.com/cdn-cgi/image/format=auto,width=720,fit=scale-down,onerror=redirect/uploads/asset/file/8db156ce-1366-442a-b319-08d127f9a8fa/Copie_de_ShanghAI_Focus.png?t=1790093611)
CTO FAQ
What engineering leaders ask us
Can open-weight coding agents replace Claude Code?
For routine work, yes. In our test, Qwen matched Claude on about 85% of daily engineering tasks (scaffolding, schema updates, tests, CLI debugging). Claude kept the lead on complex multi-file refactors with hidden edge cases.
What hardware do I need to self-host?
Qwen3.8-27B runs on a single GPU, or a laptop with quantization. Frontier mixture-of-experts models (GLM-5.3 at 744B, Qwen3.8-Max at 2.4T, Kimi K3 at 2.8T) need multi-GPU nodes, so most teams start on a hosted endpoint.
How do I keep proprietary code private?
Self-host the weights, or route sensitive prompts to a local model and send the rest to a cheap API. A coding agent on Qwen can run fully air-gapped with zero data egress.
Which licenses should legal check?
Qwen3.8-27B is Apache 2.0 and GLM-5.3-Flash is MIT. Most flagships, including GLM-5.3, Qwen3.8-Max and Kimi K3, use custom licenses with extra conditions. Read the license file before you ship a hosted product.
The CTO briefing, every Monday.
Free: benchmarks, deployment playbooks and the open models worth switching to.
Ready to move your stack? LeanAI audits your pipelines and deploys open-weight runtimes.
Book an architecture audit