If the Big Labs Block You: Every Alternative That Trains Its Own Models
To appear below, a company must train its own models, let you sign up and pay today with no sales call, and not be one of the giants. Out of scope: OpenAI, Anthropic, xAI, Google, Microsoft, Meta, Amazon and Apple — those are the other article — plus all Chinese developers. Chinese services get their own piece. Prices as of early October 2026.
The list
| Developer | Country | Type | Service | Paid plans | Where it falls short |
|---|---|---|---|---|---|
| Mistral AI | 🇫🇷 France | Text, code | Vibe (was Le Chat) | Free · Pro €14.99 · Team €24.99/user | Reasoning is the weak axis — ~43.9% GPQA Diamond, 30+ points behind frontier reasoning. 128K context; “unlimited” Pro has a soft cap near 120–150 messages. Review |
| Cognition | 🇺🇸 USA | Code | Windsurf + Cascade | Free · Pro $20 · Teams $40/user · Max $200 | A coding tool, not a general assistant. And Cognition won’t say which open base SWE-1.5 started from, so lineage is unverifiable. Announcement |
| Sarvam AI | 🇮🇳 India | Text | Indus | Free only | 105B across 22 Indian languages — strong at Indic work, behind leaders at English reasoning, and not built for code. No paid tier to escalate to. Review |
| Midjourney | 🇺🇸 USA | Image | midjourney.com | Basic $10 · Standard $30 · Pro $60 · Mega $120 | Still the benchmark for artistic taste, but no free tier and no trial — $10 minimum to see anything. Weaker than rivals at rendering text in images. Pricing |
| Black Forest Labs | 🇩🇪 Germany | Image | bfl.ai | Credit packs: 100 images $3 → 5K $150 | The photorealism leader and genuinely open (FLUX.2 klein 4B is Apache 2.0). Weakness is product, not model: credit packs rather than a polished consumer subscription. |
| Stability AI | 🇬🇧 UK | Image | stability.ai | 100 images $6.50 → 5K $325 | Open weights and self-hostable, but roughly 2× Black Forest Labs’ price per image and no longer the quality leader it was. |
| Ideogram | 🇺🇸 USA | Image | ideogram.ai | Free tier · API 100 images $6 → 5K $300 | The one model that reliably renders readable text in images — that’s the reason to use it. General aesthetic quality trails Midjourney and FLUX. |
| Recraft | 🇺🇸 USA | Image | recraft.ai | Free to start · from $10/mo | Outputs editable vectors, which nothing else here does — the pick for design assets. Narrow: not a general-purpose image model. |
| Krea | 🇺🇸 USA | Image, video | krea.ai | Basic $9 · Pro $35 · Max $105 | Real-time canvas and a good front end, but it’s mostly an aggregator over 60+ models — its own training is limited, so you inherit other vendors’ availability. Pricing |
| Runway | 🇺🇸 USA | Video | runwayml.com | Standard $15 · Pro $28 · Unlimited $95 | The most granular creative control in video, and the editor professionals actually use. Raw output quality now trails the Chinese leaders (Kling v3) and Google’s Veo. |
| Luma AI | 🇺🇸 USA | Video | lumalabs.ai | Lite $9.99 · Pro $29.99 · Unlimited $94.99 | Cheap entry and fast iteration. Weaker physics and prompt adherence than the current leaders; credit maths gets expensive at high resolution. Pricing |
| Pika | 🇺🇸 USA | Video | pika.art | from ~$8 · mid-tier ~$35 | Cheapest way into AI video and good at short stylised clips. Shortest clip lengths and the weakest of the three at realism. |
Twelve companies. Nine of them are American, which is worth sitting with: remove the giants and the Chinese labs, and the independent market is still mostly US — one French company, one Indian, one German, one British.
Where the gaps are
- Text and code is almost empty. Mistral is the only independent generalist with a real subscription. Cognition covers code only. Sarvam is free, regional and not a dev tool. That’s the whole list.
- Images are healthy. Six genuine options at real prices, with distinct strengths — taste (Midjourney), photorealism (FLUX), text rendering (Ideogram), vectors (Recraft). This is the one category where being blocked costs you nothing.
- Video is adequate but no longer leading. Runway, Luma and Pika all work and all cost under $30 to start, but the quality frontier has moved to Kling and Veo — both outside this list.
Who’s missing, and why
- Shut down: Naver’s CLOVA X closed 9 April 2026; AI21 halted Jamba development in May 2026 after cutting 61% of staff.
- No self-serve signup: Cohere (North has no public price), LG AI Research (ChatExaone needs corporate email), Upstage, Reka, NVIDIA, IBM — own models, but enterprise or API only.
- Japan has a real ecosystem, and none of it qualifies — worth spelling out, since it’s the obvious question. Sakana AI runs Sakana Chat, a free consumer chatbot, but it’s geo-restricted to Japan, and its Namazu models are post-trained on Kimi K2.6 and DeepSeek V3.1 — so it fails on access and lineage. Preferred Networks’ PLaMo 3.0 Prime is trained from scratch but keeps weights private and sells API access only. NTT’s tsuzumi and SoftBank/SB Intuitions’ Sarashina are enterprise. Fujitsu’s Takane is Cohere-based. And Rakuten AI 3.0 is a ~700B MoE under Apache 2.0 — genuinely open, but there’s no subscription to buy, so it belongs in the downloadable survey rather than here.
- No service to subscribe to: Ai2 (OLMo), EleutherAI, OpenGPT-X/Fraunhofer (Teuken), and Switzerland’s Apertus — all publish weights, none sell a product. Apertus in particular is reachable only through a partner platform, which makes it self-hosting.
- Wrong lineage: Cursor’s Composer 2 is built on Kimi K2.5, a Chinese open model, with about three quarters of the compute Cursor’s own. Writer’s Palmyra X6 is reportedly post-trained GLM-5.2, also Chinese. Perplexity’s Sonar is fine-tuned Llama. All excluded on provenance rather than quality.
That last group is the uncomfortable one. Two prominent “Western” coding products run on Chinese open weights — disclosed by Cursor, undisclosed by Cognition, who are on this list. Vendor nationality does not establish model lineage. If that distinction matters to you, ask what the base model was, and accept that sometimes the answer is withheld.
My take
For images, being cut off is a non-event — pick from six. For video you’ll manage at a quality cost. For text and code, the independent market is one French company at €14.99. That isn’t a diversified supply chain.
So the honest answer to “what if they block me” is still the one nobody markets: hold the weights. See the multi-GPU survey for what’s downloadable and the Strix Halo loadout for running it on one machine. After two shutdowns inside six months among the names that would have been on this list a year ago, that stops reading as paranoia and starts reading as planning.