{"slug": "the-last-vendor", "title": "The Last Vendor", "summary": "OpenAI's GPT-5.2, released in August 2026, achieves 95.4% on SWE-bench Verified, surpassing Claude Opus 5 by 0.6 points and becoming the top open-weights model, according to the Vals AI leaderboard. The model is available via OpenCode Go's flat-rate subscription with zero data retention, positioning it as a viable alternative to closed frontier models for developers prioritizing privacy and cost predictability.", "body_md": "# The Last Vendor\n\nOpen weights got close enough to the frontier that the vendor stopped mattering.\n\nAfter my [farewell to the frontier labs](/adieu-claude/), I still had to figure out whose GPUs I would rent for inference instead.\nMy requirements were clear:\n\n- Near-frontier capability.\n- Zero data retention (ZDR).\n- Hosting in the EU or US.\n- 256k+ context windows.\n- Native OpenCode integration.\n- A flat, predictable subscription, no nasty pay-as-you-go (PAYG) surprises.\n\nThe [SWE-bench Verified leaderboard](https://www.vals.ai/benchmarks/swebench) of August 2026 shows that 7 out of 87 models reach 95% or better.\nTwo of those seven publish their weights: DeepSeek V4 Pro and GLM-5.3.\nDeepSeek V4 Pro sits second overall, only 0.6 percentage points behind Claude Opus 5 and ahead of every other closed model.\n\nOpen weights are a different commodity than pure open source: [OLMo](https://allenai.org/olmo) publishes its weights, its training data, and its code, so it can be reproduced, audited, and built on without asking anyone’s permission.\nUnfortunately, OLMo is nowhere near the frontier.\nThe practical choice is not between open and closed but between open weights and closed.\nAt least the weights escape the vendor, and that breaks the lock-in.\n\nAdvertised context windows are a gimmick: Google’s own [model card](https://deepmind.google/models/model-cards/gemini-3-1-pro/) 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.\nThe 1M windows on the open models are real but unvalidated, so they are best seen as headroom, not a guarantee.\n\n##\nThe menu\n\nI verified every candidate against vendor documentation and live API endpoints in August 2026.\nNot included below are [Z.ai](https://docs.z.ai/devpack/overview) and [Kimi](https://www.kimi.com/) subscriptions, because they are yet again single vendors that also happen to run on Chinese infrastructure, which I prefer to avoid.\n[Hugging Face PRO](https://huggingface.co/pricing) is out too, because it routes to various third-party providers: you inherit their retention and region quirks.\n\nWhat remains:\n\n| Provider | Price | Top models (SWE-bench · context) | Retention | Hosted in | Verdict |\n|---|---|---|---|---|---|\n|\n\nKimi K3 · 93.40% · 1M\n\nHy4 preview · 80% (10-task) · 1M\n\n[zero](https://opencode.ai/docs/go/)on the models I use**best fit**[Synthetic](https://synthetic.new/?referral=zLvKGKYY)GLM-5.2 · 82.80% · 512K\n\n[zero](https://synthetic.new/policies/privacy)**overflow**[Solheim](https://solheim.ai/)[Melious](https://melious.ai/pricing)[Venice](https://venice.ai/pricing)[no logs by default](https://venice.ai/privacy), up to[E2EE](https://docs.venice.ai/guides/features/tee-e2ee-models)[Chutes](https://chutes.ai/pricing)[no logging](https://chutes.ai/privacy)[Featherless](https://featherless.ai/)[no prompt logging](https://featherless.ai/docs/privacy-and-logging)Melious only lists EU countries upon registration, which effectively means it is not available to companies and people outside.\n\n##\nThe selection\n\nMy main subscription is [OpenCode Go](https://opencode.ai/go?ref=PET2T3Y3AD).\nIt 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.\nThe 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.\nThe two best open models on OpenCode Go are unreachable: `deepseek-v4-pro`\n\nand `deepseek-v4-flash`\n\nare hosted in China and require an explicit opt-in.\nThe practical leaderboard therefore tops out at GLM-5.3 (95.40%), with Kimi K3 (93.40%) behind it.\n\nFor overflow, I use [Synthetic](https://synthetic.new/?referral=zLvKGKYY).\nGo’s limits run out fast on complex tasks that require large contexts.\nSynthetic’s packs are stackable, and each pack carries the same 500 requests per five hours: you can put up to five on top of one another, and the whole arrangement resets monthly with one free reset during the month.\n\nPrompt caching on OpenCode Go requires no configuration and is effective, so you get roughly four times as much usage out of the subscription as its headline rates suggest. Caching works automatically on Synthetic too, even though its documentation never mentions the word “cache” once across its pages.\n\n##\nThe fine print\n\nGo’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](https://github.com/anomalyco/opencode/commit/2df47ee3948e).\nThe DeepSeek models are hosted in China and gated behind a China opt-in in the settings.\nThe existence of that switch is itself evidence that the rest are not hosted in China, but where their roster runs is an open question.\nDeepSeek’s ZDR agreement was [renewed monthly](https://opencode.ai/docs/go/) and expired on 31 August 2026 without ceremony.\nI therefore do not use DeepSeek’s models through OpenCode Go anymore.\n\nSynthetic is transparent: its inference servers are located in the US with GLM-5.2 also hosted from Sweden and Iceland:\n\n```\ncurl -sS https://api.synthetic.new/openai/v1/models \\\n  -H \"Authorization: Bearer $SYNTHETIC_API_KEY\" \\\n  | jq -r '.data[] | [.id, ([.datacenters[].country_code] | join(\",\"))] | @tsv'\n```\n\n##\nThe last vendor\n\nMy setup—the skills, the prompts, the config—are plain text on open standards, untethered to any model or any provider.\nThe licence matters: GLM-5.2 under MIT is a safer long-term bet than GLM-5.3 under a [custom licence](https://huggingface.co/zai-org/GLM-5.3/raw/main/LICENSE) with a $10bn revenue trigger and a security review whose scope Z.ai itself determines.\nAnd when I swap GLM-5.3 for whatever comes next, I only have to change one line of configuration.\n\nThe 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.", "url": "https://wpnews.pro/news/the-last-vendor", "canonical_source": "https://ianreppel.org/the-last-vendor/", "published_at": "2026-08-31 22:00:00+00:00", "updated_at": "2026-09-01 02:52:19.095535+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-products", "ai-tools"], "entities": ["OpenAI", "GPT-5.2", "Claude Opus 5", "SWE-bench Verified", "Vals AI", "OpenCode Go"], "alternates": {"html": "https://wpnews.pro/news/the-last-vendor", "markdown": "https://wpnews.pro/news/the-last-vendor.md", "text": "https://wpnews.pro/news/the-last-vendor.txt", "jsonld": "https://wpnews.pro/news/the-last-vendor.jsonld"}}