Together adds NVIDIA B200 post-training at $11/GPU/hr, $9 launch promo
By Rob Litterst, Co-founder, PricingSaaS
Published
Together’s pricing page · Change detected
What changed
Together added a new Post-Training product to its pricing page: on-demand training infrastructure on NVIDIA HGX B200 GPUs, billed per minute, at $11.00/GPU/hr list with a launch promo price of $9.00/GPU/hr valid until 12/14/2026. In the same update, Together added DeepSeek V4.1 Flash 1017 to its serverless inference pricing at $0.30 input, $0.006 cached, $1.20 output per 1M tokens.
Why it matters
The B200 post-training launch extends Together's infrastructure stack beyond serverless inference and fine-tuning into on-demand training compute, competing more directly with dedicated GPU-cloud providers on raw hardware access rather than managed model endpoints. Watch whether the $9.00 launch promo reverts to the $11.00 list price when it expires 12/14/2026, and whether Together keeps layering new hardware SKUs onto this page at its current pace.
In context
Together's pricing page rarely sits still: this B200 post-training launch lands one day after the company ended a free Muse Spark Open promo, and follows a run of serverless-rate tweaks — Reflection Beam added Oct 17, Kimi K3 cached pricing raised Oct 14, and Kimi K3 cut 10% while Qwen3.8 Flash rose Oct 6. Together averages a change about every 8 days; across the index, Pika, Ideogram and Mintlify make credits and usage-billing changes most often. Launching new hardware at a discount, then presumably reverting to list, is a familiar way to seed usage on new infrastructure before committing to it at full price.
What people are saying
“That is Great deal: - you own the training loop, data, rewards, custom losses. We handle the GPUs, tuned rollout engines, and weight sync. - RL stays consistent between training and serving. Multiple LoRAs run in parallel on the same GPUs. - Train, compare adapters, deploy the winner, or download the weights. I’d take that deal 10 times 😀”
View original“I like how Together AI is literally getting into all aspects of AI development. With the newly launched post-training service. they’re truly covering the field, proving that focus is indeed not chic in this era”
View original“SFT-as-a-service is a hosting problem. RL-as-a-service is a verifier problem. Once users define their own rewards, your job becomes stopping reward hacking at scale, not running rollouts. What guardrails do you ship: verifier templates, sandboxes, or caveat emptor?”
View original