The Last Vendor

Open weights got close enough to the frontier that the vendor stopped mattering.

After my farewell to the frontier labs, I still had to figure out whose GPUs I would rent for inference instead. My requirements were clear:

• Near-frontier capability.
• Zero data retention (ZDR).
• Hosting in the EU or US.
• 256k+ context windows.
• Native OpenCode integration.
• A flat, predictable subscription, no nasty pay-as-you-go (PAYG) surprises.

The SWE-bench Verified leaderboard of August 2026 shows that 7 out of 87 models reach 95% or better. Two of those seven publish their weights: DeepSeek V4 Pro and GLM-5.3. DeepSeek V4 Pro sits second overall, only 0.6 percentage points behind Claude Opus 5 and ahead of every other closed model.

Open weights are a different commodity from pure open source: OLMo publishes its weights, its training data, and its code, so it can be reproduced, audited, and built on without asking anyone’s permission. Unfortunately, OLMo is nowhere near the frontier. The practical choice is not between open and closed but between open weights and closed. At least the weights escape the vendor, and that breaks the lock-in.

Advertised context windows are a gimmick: Google’s own model card reports Gemini 3.1 Pro at 84.9% on MRCR at 128k and 26.3% at 1M, a 58.6-point drop buried in the evaluation table while the 1M window leads the marketing. The 1M windows on the open models are real but unvalidated, so they are best seen as headroom, not a guarantee.

The menu

I verified every candidate against vendor documentation and live API endpoints in August 2026. Not included below are Z.ai and Kimi subscriptions, because they are yet again single vendors that also happen to run on Chinese infrastructure, which I prefer to avoid. Hugging Face PRO is out too, because it routes to various third-party providers: you inherit their retention and region quirks.

What remains:

Melious only lists EU countries upon registration, which effectively means it is not available to companies and people outside the EU.

OpenCode Go publishes nothing on concurrency limits. Synthetic allows one concurrent request per model per pack (stackable) and queues the rest rather than dropping them. Ollama Pro supports up to three concurrent requests.

What sets Synthetic apart is that it regenerates the quotas in small increments, so you never have to wait for a full reset when you exhaust them.

The selection

My main subscription is OpenCode Go. It is the only option that meets almost every requirement at once: 1M context, zero retention on the models I use, native OpenCode integration, and a flat fee. The model list is curated rather than exhaustive: GLM-5.3, Kimi K3, GLM-5.2, Hy4 preview (a.k.a. Ox Alpha), and a long tail of cheaper models for the work that does not need thinking. The two best open models on OpenCode Go are unreachable: deepseek-v4-pro and deepseek-v4-flash are hosted in China and require an explicit opt-in. The practical leaderboard therefore tops out at GLM-5.3 (95.40%), with Kimi K3 (93.40%) behind it.

Prompt caching on OpenCode Go requires no configuration and is effective. Caching appears to work automatically on Synthetic too. Ollama has an open feature request that claims no cache in agentic loops.

For overflow, I use Synthetic. Go’s limits run out fast on complex tasks that require large contexts. Synthetic’s packs are stackable, and each pack carries the same 500 requests per five hours: you can stack up to five, and the whole arrangement resets monthly with one free reset during the month. Compared to Synthetic, Ollama’s larger selection of near-frontier models makes it an attractive backup, but Synthetic’s pricing model stretches further for my usage.

A typical month as derived from opencode stats across both personal and work machines uses: 1.2B input/context tokens, 6.7M output tokens, 17,000 messages, with a 94% cache-read rate. At public rates, that mix on GLM-5.3 would cost roughly $430 a month. Flat subscriptions change the arithmetic. OpenCode Go is the baseline: $10 a month buys $60 of usage credits, and when the limit runs out I use its free models, such as Big Pickle. Synthetic publishes the equivalent too: each $30 pack carries $24 a week of credits, about $102 a month of value. Ollama Pro’s $20 buys $60 of credits. None of the flat fees alone covers a full month of frontier-model usage at my volume, which is why the frontier work and the overflow go to Synthetic.

The fine print

Go’s hosting location is unverified: the docs claimed “US, EU, and Singapore” until 31 July 2026 when that fragment was deleted from all locale files. The DeepSeek models are hosted in China and gated behind a China opt-in in the settings. The existence of that switch is itself evidence that the rest are not hosted in China, but where their roster runs is an open question. DeepSeek’s ZDR agreement was renewed monthly and expired on 31 August 2026 without ceremony. I therefore do not use DeepSeek’s models through OpenCode Go anymore.

Synthetic is transparent: its inference servers are located in the US with GLM-5.2 also hosted from Sweden and Iceland:

curl -sS https://api.synthetic.new/openai/v1/models \
  -H "Authorization: Bearer $SYNTHETIC_API_KEY" \
  | jq -r '.data[] | [.id, ([.datacenters[].country_code] | join(","))] | @tsv'

Ollama offers frontier DeepSeek models hosted in the US and Europe.

The last vendor

My setup—the skills, the prompts, the config—is untethered to any model or any provider. The licence matters: GLM-5.2 under MIT is a safer long-term bet than GLM-5.3 under a custom licence with a $10bn revenue trigger and a security review whose scope Z.ai itself determines. And when I swap GLM-5.3 for whatever comes next, I only have to change one line of configuration.

The models are close enough to the frontier that the gap is a fraction of a point, and for most of what a coding agent does, a fraction of a point is not a reason to stay. It is the last time I need to care, because I can leave.