What changed
Fireworks' current pricing page publishes a two-column transition for its on-demand GPU deployments: prices through August 31 and new rates from September 1, 2026. H100 80GB and H200 141GB increase from $7 to $8 per GPU-hour; B200 180GB rises from $10 to $13; B300 288GB from $12 to $15; and GB300 288GB from $18 to $20. Fireworks bills on-demand deployments per GPU second with no startup charge, and its Dedicated Training API uses the same GPU-hour pricing, so the change affects both dedicated serving and GPU-hour training jobs. Region-restricted deployments continue to carry a 1.5× premium.
Why it matters
Dedicated GPU economics can dominate the unit cost of high-volume inference, fine-tuning and reinforcement-learning workloads. These increases range from roughly 11% to 30%, large enough to change break-even points between accelerator generations, serverless APIs and competing GPU clouds. The biggest increase is on B200, where a workload consuming 1,000 GPU-hours per month moves from $10,000 to $13,000 before regional premiums or other services. Teams should therefore benchmark completed-task throughput rather than choosing hardware on hourly price or peak specifications alone.
The September increase varies materially by accelerator
H100 and H200 move from $7 to $8 per hour, about a 14.3% increase. B200 rises 30% from $10 to $13, B300 rises 25% from $12 to $15 and GB300 rises about 11.1% from $18 to $20. The uneven changes can alter the relative cost ranking of GPUs even when performance remains unchanged.
Training inherits the same GPU-hour table
Fireworks states that Dedicated Training API jobs are priced per GPU hour using the On-Demand Pricing section. Teams running reinforcement fine-tuning or dedicated training should therefore update both serving and training forecasts, not only inference deployment budgets.
Per-second billing limits idle waste but not rate exposure
On-demand deployments are billed by the second and do not charge extra startup time, which remains useful for bursty workloads. But sustained fleets still absorb the full hourly increase, and a region-restricted deployment costs 1.5 times the listed rate.
Hourly price is not total inference cost
A more expensive accelerator can still win if it materially improves throughput, latency or batch efficiency. Builders should compare cost per successful request, training run or completed agent task under the exact model, quantization, context and concurrency they use.