Together AI Pricing in 2026: Every Serverless Model Priced

Together AI's serverless card carries 25 open-weight chat models as of 2026-11-10, running from $0.15/$0.50 per million tokens (GLM-5.3 Flash) to $3.00/$15.00 (Kimi K3) — with the deepest cache discounts in the series so far (98% off on DeepSeek V4.1 Flash) and a Batch API that bills at half price.

Key facts

• The serverless card is 25 chat rows, priced input/cached-input/output: Kimi K3 at $3.00 / $0.30 / $15.00, Kimi K3.1 Pro at $2.00 / $0.20 / $10.00, GLM-5.3 at $1.40 / $0.26 / $4.40, Muse Spark Open 816B A42B at $1.25 / $0.15 / $4.25, DeepSeek V4 Pro 0813 at $1.32 / $0.13 / $3.96, DeepSeek V4.1 Flash at $0.30 / $0.006 / $1.20, and GLM-5.3 Flash at $0.15 / $0.03 / $0.50 (Together AI pricing, accessed 2026-11-10).
• Cache discounts run 80–98% off: DeepSeek V4.1 Flash caches input at $0.006 versus $0.30 (98% off), MiniMax M3.1 Flash at $0.015 (90% off), GLM-5.3 at $0.26 (81% off) — with no separate cache-write fee published (Together AI pricing).
• Batch API is the 50% lever: selected serverless models run at up to 50% off serverless rates in Batch mode, with a separate rate-limit pool, in exchange for a job-shaped (not request-shaped) workflow (Batch API docs).
• No numeric rate-limit table exists: serverless is best-effort — the docs say you may hit 429/503 under high demand and should back off exponentially; guaranteed capacity is sold as Provisioned Throughput ($0.05 per PTU/minute on the pricing estimator) or dedicated endpoints (Rate limits docs; Together AI pricing).
• Prompt storage is on by default: Together stores prompts and responses "and may use them for product improvements" unless an organization enables zero data retention; training on customer data is opt-in and off by default (ZDR docs; Privacy policy).
• Footing: Together AI raised $800 million at an $8.3 billion valuation in July 2026, led by Aramco Ventures, with annual bookings it said crossed $1.15 billion the prior quarter (Reuters, 2026-07-01).

What Together AI is

Together AI (Together Computer, Inc., founded 2022, San Francisco) is an inference and fine-tuning cloud built around open-weight models, reached through an OpenAI-compatible API. The company sells four deployment tiers: serverless per-token inference, Provisioned Throughput (reserved token capacity priced in PTUs), Dedicated Model Inference (single-tenant reserved GPUs), and raw GPU clusters by the hour. The catalog spans 200+ models across chat, image, video, audio, and embeddings; this page prices the serverless chat card, which is what most builders touch. Together also builds its own experimental models — the card carries Tev1 4B (input $0.04, output free) and hosts Reflection's Beam 1.0 and Inkling from Thinking Machines.

The company's footing, from primary filings and coverage: a $305 million Series B at $3.3 billion in February 2025, then an $800 million Series C at $8.3 billion on July 1, 2026, led by Aramco Ventures with NVIDIA, Salesforce Ventures, General Catalyst, Vista Equity Partners, and others participating (Reuters). At that raise the company said annual bookings had crossed $1.15 billion the prior quarter, and that it expects compute capacity to expand roughly 50-fold over five years.

The full serverless card, priced (2026-11-10)

Every open-weight chat model on the serverless card, first-party list rates accessed 2026-11-10:

Two structural notes. First, the Flash tiers own the price-performance end: GLM-5.3 Flash, Qwen3.8 Flash, MiniMax M3.1 Flash, and DeepSeek V4.1 Flash all sit at or below $0.30 input, and the same models carry the deep cache discounts — these are the distilled routers Together itself benchmarks for cascade deployments. Second, premium pricing tracks recency and scale, not intelligence alone: Kimi K3.1 Pro ($2.00/$10.00) undercuts its own older sibling Kimi K3 ($3.00/$15.00) while ranking higher on third-party intelligence indices, a pricing inversion worth checking before defaulting to the older row.

Cache economics and the Batch lever

Together's cached-input schedule is aggressive on exactly the models agentic workloads hammer. DeepSeek V4.1 Flash caches input at $0.006 per million tokens — 98% off the $0.30 list and, per DeAI's cross-host tracking, matched only by DeepInfra and Fireworks among the hosts priced this quarter. MiniMax M3.1 Flash caches at 90% off, GLM-5.3 at 81% off, Kimi K3 at 90% off. The published card shows no separate cache-write fee.

On top of caching, the Batch API bills selected serverless models at up to 50% off for job-shaped workloads — you submit a file of requests and poll for results rather than holding a connection open — and batch jobs get their own rate-limit pool, so they don't consume your interactive headroom (Batch API docs). Stacked: DeepSeek V4.1 Flash under Batch is roughly $0.15 input / $0.60 output, with cached input at $0.003. For anything latency-tolerant — evaluation runs, backfills, bulk classification — that combination is the cheapest published way to run a frontier-grade open model in this series so far.

Worked example, assumptions labeled: an agentic request with a 20k-token system prompt (95% cache hit) and a 2k-token response on DeepSeek V4.1 Flash bills about $0.0003 uncached input + $0.000114 cached + $0.0024 output ≈ $0.0028 per request serverless, or ~$0.0014 under Batch. Your own hit rate decides how much of that transfers; the estimator on the pricing page is the right way to price your profile.

Speed: third-party numbers

Artificial Analysis benchmarks 18 Together-hosted endpoints continuously (every 6 hours, 5-day averages). Its highlights, accessed 2026-11-10: on gpt-oss-120b, Together's endpoint shows an error rate of 7.0–8.3% and a 1.81–2.07s time to first answer token on the low-reasoning variant — well off the pace of the custom-silicon hosts DeAI tracks, where Cerebras and Groq hold the throughput lead on the same model. On intelligence, Together's catalog tops out at Kimi K3.1 Pro (index 46), DeepSeek V4.1 Flash 1017 (33), and GLM-5.3 (32), and its cheapest blended prices are $0.09–0.10/M on Nemotron 3.5 Super VL, MiniMax M3.1 Flash, and GLM-5.3 Flash (Artificial Analysis). Together is selling breadth and fine-tuning depth, not raw tokens-per-second; if maximum throughput is the requirement, the fast-silicon specialists remain the alternatives — our Together AI alternatives page maps that field.

Privacy: storage on by default, training off

The distinction that matters on Together is storage versus training. The privacy policy (updated December 17, 2025) states plainly that Together does not train on customer data without explicit opt-in — and that opt-in is off by default (Privacy policy). But unless your organization separately enables zero data retention in Privacy settings, Together stores the prompts you send and the responses models return and may use them for product improvements. ZDR is self-serve (one toggle), applies only to requests after it's enabled, and — critically — disables passthrough models (Together-hosted gateways to third-party providers like OpenAI and Google), because passthrough requires retained prompts. Files uploaded for fine-tuning or batch jobs remain stored until you delete them. DeAI treats all of these as self-published policy statements, not independently verified attestations.

Rate limits, exactly

Together publishes no numeric TPM/RPM table — a genuine difference from Fireworks and Groq, whose adaptive-limit tables we've priced. The docs describe serverless as best-effort: under high demand you may see 429 Too Many Requests or 503 Service Unavailable, the fix is exponential backoff, and model-specific limits can appear on high-demand models (Rate limits docs). Guaranteed capacity comes in two forms: Provisioned Throughput, reserved token capacity priced in PTUs at $0.05 per PTU-minute on the published estimator (throughput per PTU varies by model and token type), and Dedicated endpoints on reserved GPUs — H100 on-demand at $5.49/GPU-hour, B200 at $8.99. Raw GPU clusters run $3.99 (H100) to $9.99 (B300) per hour on demand, with preemptible capacity from $1.99.

How to choose

Together is the fit when you want the broadest open-model catalog behind one OpenAI-compatible endpoint, fine-tuning in the same bill, and deep cache discounts on Flash-tier models — or when you need dedicated capacity without moving vendors. It is not the fit when a published, numeric rate-limit table is a hard requirement (Fireworks and Groq publish one), when maximum tokens-per-second is the goal (Cerebras, SambaNova, Groq), or when the absolute floor price is the only criterion (DeepInfra and aggregator listings hold that end). Prices move — DeAI's weekly price-index deltas have recorded repeated repricings across this field — so treat every figure here as dated to 2026-11-10 and re-verify against Together's pricing page before committing spend.