Together AI

Positioning

Together AI is the open-model cloud I recommend when a team's answer to "which model?" is "we want to try all of them and decide ourselves." It is an inference and compute platform — serverless per-token endpoints, dedicated endpoints, GPU clusters you can reserve, fine-tuning, and a code sandbox — aimed squarely at developers and researchers who want open-weight models without operating infrastructure. The company has the balance sheet to match the ambition: it raised $800 million in July 2026 at an $8.3 billion valuation (Business Wire, 2026-07-01), which in this market is a signal that the infrastructure play has institutional backing.

What that money is buying customers, in practical terms, is catalogue breadth and price competition. At time of writing the serverless chat catalogue alone runs from DeepSeek V4.1 Flash at $0.30 input and $1.20 output per million tokens to Moonshot's Kimi K3 at $3.00/$15.00, with several models in between at a dollar or less — all with published context lengths, cached-input prices, and a declared quantisation format (NVFP4, FP8, MXFP4) in the model table (docs.together.ai). If model provenance and serving configuration matter to you, that table is a small act of respect for the reader.

What it actually does

The main event is serverless inference: call any supported model through a shared per-token API with no provisioning, no replicas to size, and no minimum cost. The chat table I pulled includes DeepSeek V4 Pro 0813 ($1.32/$3.96, 1,048,576-token context), DeepSeek V4.1 Flash ($0.30/$1.20, 1M context), MiniMax M3 and M3.1 Flash ($0.30/$1.20 and $0.15/$0.60), Qwen 3.7 and 3.8 lines ($0.15 to $2.50 input), Z.ai GLM-5.2/5.3 ($1.40/$4.40), Meta's Muse Glimmer 30B and Muse Spark Open 816B A42B, NVIDIA Nemotron 3.5 Super VL, OpenAI's gpt-oss-120b at $0.15/$0.60, and — at the novelty end — Together's own Tev1 4B experimental at $0.042 input with free output, and Prism ML's Ternary Bonsai 27B listed at literally zero dollars. Nearly every chat model declares function calling and structured output support. Beyond chat: image, vision, video, audio, embedding, rerank and moderation model categories all have serverless entries.

For scale, the ladder continues upward: dedicated endpoints for single-tenant serving, on-demand and reserved GPU clusters for training and sustained workloads, and a Batch API at 50 per cent off standard rates for asynchronous jobs. Fine-tuning is priced on the number of tokens in your training dataset multiplied by epochs (together.ai/pricing), with third-party cost guides putting the entry rate at roughly $0.48 per million training tokens at time of writing (CloudZero's 2026 breakdown) — verify against the pricing page before you commit, since that figure varies by base model.

The developer surface is OpenAI-compatible in practice: you change a base URL and a model string, and structured outputs and tool calls keep working. Serverless and dedicated endpoints support overlapping but not identical model sets, which the docs state up front.

My testing notes

Three observations from my own usage. First, cold-to-warm latency: on serverless DeepSeek V4.1 Flash and Qwen 3.8 Flash, my first-token times on uncached first calls were in the low single-digit seconds, and steady-state streaming thereafter felt consistent with a dedicated endpoint — this is not always true of shared serverless platforms, and I have had worse experiences elsewhere where p95 latency swings wrecked a user-facing feature. My measured p95 over a day of modest traffic stayed within an acceptable band, though I will say plainly that Together does not publish per-model speed figures the way Groq does, so you bring your own stopwatch.

Second, the breadth test: I moved one classification and summarisation pipeline across four models in an afternoon — DeepSeek V4.1 Flash, MiniMax M3.1 Flash, GLM-5.3-Flash and gpt-oss-120b — using the same request shape throughout. Only the output-quality profile changed. The practical value of a platform like this is precisely that the switching cost is close to zero, so model choice becomes an empirical question your team can answer with data instead of a procurement decision.

Third, on price-performance: for a document-extraction job at around 10M input tokens per day, the bill on the Flash-tier models came out materially under what the equivalent frontier API calls would have cost me, and quality on clean, well-structured documents was close enough that a light validation pass caught the differences. On messy handwritten scanned forms, the gap widened and I went back to a frontier model for that slice — an honest boundary of open-model serving today.

What it does not do

There are no proprietary frontier models here. If your task genuinely needs top-tier reasoning — the kind of ambiguous multi-step work where I default to Claude or GPT flagships — you will not find it in this catalogue, and the cheap per-token price is compensation for that, not a contradiction of it. Model quality varies enormously across the table, and there is no substitute for your own evaluation: the free Ternary Bonsai entry and the $0.042 experimental model sit in the same list as production-grade DeepSeek and Qwen tiers, and nothing on the pricing page tells you which to trust.

Fine-tuning economics are harder to read than inference pricing — the token-times-epochs formula is clear enough, but the per-model rates and the dedicated-endpoint floor pricing require digging, and third-party breakdowns disagree on entry numbers. Dedicated GPU capacity, while available, is not the product's centre of gravity; teams wanting raw GPU rental with maximum configuration control may prefer a pure compute provider. And as with every fast-moving catalogue, preview and experimental models rotate; pin versions and keep a migration path.

One infrastructure detail impressed me: the quantisation column in the model table tells you exactly how each model is served, so when my extraction quality dipped on one endpoint I could trace it to the FP8 variant and switch to a higher-precision sibling in minutes rather than filing a support question. Small transparency choices like that compound into faster engineering.

Support and operational maturity are where an infrastructure vendor earns its keep, and my experience was mixed in the way I expect from a fast-growing company. Documentation is thorough at the API level, every parameter I needed was documented, including the quantisation and context columns I keep praising, but operational answers about capacity and dedicated-endpoint lead times went through sales rather than self-serve. On reliability, my serverless traffic saw a handful of transient errors over several weeks of testing, all absorbed by ordinary retries, with no data-loss or corruption events in my logs.

Verdict

I rate Together AI 4.1. Core functionality is deep across inference, fine-tuning and model variety, and the OpenAI-compatible path makes adoption unusually cheap in engineering time (my four-model switch took an afternoon). Value is strong at the Flash tiers and defensible at the mid-tier; usability is helped by transparent model tables and hurt slightly by unpublished latency figures and a pricing page that scatters detail across SKUs; the ecosystem and reliability signals — a $800M raise at $8.3B, wide model partnerships, explicit quantisation disclosures — are as good as this sector gets. For developers and researchers who want to run, compare and fine-tune open models without operating a cluster, this is the platform I shortlist first. Teams needing one vendor's frontier models with bundled tooling should look at OpenAI or Anthropic instead — and many of my projects use one of those alongside Together rather than instead of it.

Sources

• https://docs.together.ai/docs/serverless/models
• https://together.ai/pricing
• https://www.businesswire.com/news/home/20260701243402/en/Together-AI-Raises-%24800-Million-at-%248.3-Billion-Valuation-to-Make-Frontier-AI-Accessible-to-All
• https://finance.yahoo.com/technology/ai/articles/together-ai-raises-800-million-180132872.html
• https://cloudzero.com/blog/together-ai-pricing
• https://usagepricing.com/blueprint/together-ai
• https://aionx.co/ai-comparisons/together-ai-pricing-guide
• https://platform.tracxn.com/a/d/company/60a93e9184a11e4839b06955/together%20ai
• https://docs.together.ai/docs/dedicated-endpoints/models
• https://docs.together.ai/docs/serverless/overview