2027 planning guide — evidence checked 20 November 2026

Fireworks vs Together AI: price, throughput and dedicated GPUs

Both sell serverless inference for open-weight models, both publish OpenAI-compatible endpoints, and both will sell you dedicated GPUs. On the models they share, the sticker prices are mostly identical; the difference is in billing modes, throughput tiers and what happens when you outgrow serverless.

Short answer. On eight shared models the list prices are identical, so choose Fireworks for latency tiers, Batch and Spot discounts, and reserved throughput with an SLA; choose Together for lower published dedicated GPU rates and a broader cluster offering. The MiniMax row is not like-for-like: Fireworks lists M3 and Together lists M3.1 Flash, which are different model versions.

Fireworks and Together compete for the same open-weight workloads. Both buy the same weights, publish per-token rates and offer a path to dedicated capacity. That makes the sticker price a weak signal; cost per completed task, including cache hits, batch tolerance and idle capacity, is the better test.

How we evaluated this

We weighted five criteria for a team serving an open-weight model in production.

  • Serverless price, 30%. Input and output rates on the models both vendors list, at a stated token mix.
  • Latency control, 25%. Whether the vendor sells a faster tier and what it costs.
  • Dedicated economics, 20%. Published GPU-hour rates and cluster terms.
  • Fine-tuning fit, 15%. Billing units and minimums for a realistic dataset.
  • Operational fit, 10%. Rate limits, cold starts, regions and SDK compatibility.

We read both pricing pages on 20 November 2026. We did not run a throughput test, a cold-start test or a quality evaluation. Vendor tokens-per-second figures are self-published and describe generation speed, not end-to-end latency.

Eight shared models, one price list

Serverless list prices, USD per million tokens, checked 20 November 2026
ModelFireworks inputFireworks outputTogether inputTogether output
GLM-5.3$1.40$4.40$1.40$4.40
GLM-5.3 Flash$0.15$0.50$0.15$0.50
DeepSeek V4.1 Flash$0.30$1.20$0.30$1.20
DeepSeek V4.1 Flash 1017$0.30$1.20$0.30$1.20
Kimi K3$3.00$15.00$3.00$15.00
Kimi K3.1 Pro$2.00$10.00$2.00$10.00
Muse Spark Open$1.25$4.25$1.25$4.25
gpt-oss-120B$0.15$0.60$0.15$0.60
MiniMax M3 / M3.1 Flash$0.30$1.20$0.15$0.60
Dedicated H100$8.00 per GPU-hour$5.49 per GPU-hour
Dedicated B200$13.00 per GPU-hour$9.29 per GPU-hour

The MiniMax row is the exception that proves the rule. It is not the same model on both sides. Fireworks lists MiniMax M3 at $0.30/$1.20; Together lists MiniMax M3.1 Flash at $0.15/$0.60. A cheaper rate for a different version is not a price win. Compare the version you actually need.

Why eight prices are identical

Both providers buy the same open weights and compete on the same reference rates, so the list price converges. The bill diverges in four places: cached input, batch, spot and reserved capacity. Fireworks discounts cached input at 50% or more with no write fee; Together shows cached rates on its model table. Both publish batch pricing, but the discount varies by model.

A procurement spreadsheet with only input and output rates will show a tie. Add your cache hit rate, the share of traffic that can wait for batch, and the share that can tolerate spot interruption. Those three numbers decide the winner more often than the headline rate.

Fireworks sells modes; Together sells capacity

Fireworks publishes Standard, Priority, Fast, Ultra-fast (preview), Batch and Spot modes on the same endpoint. Fast is marketed at 100+ tokens per second; Batch is 50% off; Spot is 30% off on best-effort capacity and can drop to zero. Reserved throughput is sold in dollars per minute, with a throughput SLA and 99.9% uptime terms. Every mode is another routing decision and another line in the cost model.

Together's serverless page is simpler. It publishes per-token rates, a batch API price, provisioned throughput units and dedicated GPU instances. It does not publish one throughput figure across the catalogue on the pricing page. That is not a weakness in the product; it is a difference in what the vendor chooses to guarantee. If you need a contractual throughput number, Fireworks' reserved product is the clearer starting point. If you need lower $/GPU-hour, Together's dedicated rates are lower on the shapes we checked.

Dedicated GPU math

Assume eight H100 GPUs running 730 hours a month, no idle time, no storage and no tax. That is 5,840 GPU-hours. Fireworks lists H100 at $8.00 per GPU-hour; Together lists dedicated H100 at $5.49 and GPU-cluster H100 at $3.99 on-demand, with reserved terms from $3.45 to $3.69 depending on duration.

8 x 730 = 5,840 GPU-hours

Fireworks H100:              5,840 x $8.00 = $46,720.00
Together dedicated H100:     5,840 x $5.49 = $32,061.60
Together cluster H100:       5,840 x $3.99 = $23,301.60

Monthly gap, Fireworks vs Together dedicated: $14,658.40

At high sustained utilisation, the dedicated gap dominates the token price. At low or spiky utilisation, serverless is cheaper because you do not pay for idle GPUs. Fireworks also scales idle deployments to zero after one hour by default, a cold-start risk for low-traffic endpoints; Together describes autoscaling, so test cold starts on your own model.

Fine-tuning is billed in different units

Fireworks' Serverless Training API charges per million prefill, sample and train tokens. For GLM 5.3 it lists $4.86 prefill, $0.972 cached prefill, $12.15 sample and $14.58 train; DeepSeek V4.1 Flash 1017 lists $0.60, $0.12, $3.60 and $4.98. Dedicated training starts at $8.00 per GPU-hour.

Together charges supervised fine-tuning and DPO per million training-dataset tokens, with a minimum charge per job. Qwen3.5 27B lists $1.05 SFT and $2.62 DPO with a $4 minimum; Llama 3.3 70B lists $2.03 and $5.08 with a $4 minimum; gpt-oss-120B lists $2.50 and $6.25 with a $6 minimum. The units are not directly comparable: Fireworks separates prefill, sample and train tokens, while Together prices the training dataset. Model the same dataset and epoch count on both.

When Fireworks is the better choice

Fireworks wins when latency tiers matter, when a meaningful share of traffic can use Batch or Spot, and when reserved throughput with a 99.9% uptime term is easier to approve than a best-effort pool. It also wins when you want Fireworks Sovereign to run inference inside your own cloud account, or when its training-token billing matches how your team experiments.

When Together is the better choice

Together wins when dedicated GPU cost is the binding constraint: $5.49 versus $8.00 for H100 and $9.29 versus $13.00 for B200 is a large gap at steady utilisation. It also wins when you need GPU clusters, B300 or GB300 capacity, or a simpler fine-tuning price per dataset token. The serverless catalogue is broad enough that most open-weight production models are available on both.

When neither is the right shape

If you do not need dedicated GPUs or a provider-specific mode, a flat API is a different product with a different bill. Plugsky Builder is $60 a month with no monthly token cap and no per-token overage; Scale is $120. Fair-use concurrency and RPM limits apply. See the model catalogue and the DeepInfra comparison.

What we did not verify

We did not verify throughput, time to first token, cold-start duration, uptime or error rates. We did not test the same prompts on both. We did not verify reserved-throughput quotes, cluster availability, regional pricing or enterprise discounts. We did not test fine-tuning quality or total job time. Re-check the exact model ID before committing.

For Plugsky, we did not verify a SOC 2 certification. Request the current SOC 2 report or attestation under NDA and verify the period. Flat-plan capacity is governed by fair-use concurrency and RPM, not by an SLA credit schedule.

Frequently asked questions

Is Fireworks or Together cheaper?

On eight shared models the list prices are identical. MiniMax is not like-for-like: Fireworks lists M3 at $0.30/$1.20 and Together lists M3.1 Flash at $0.15/$0.60. Effective cost depends on cache, batch, spot and reserved capacity.

Which has faster inference?

Fireworks publishes a Fast mode at 100+ tokens per second and an Ultra-fast preview. Together does not publish one throughput figure across its catalogue. Measure time to first token and p95 on your own prompts and region.

Do both support fine-tuning?

Yes, with different billing. Fireworks charges per million prefill, sample and train tokens plus dedicated training from $8 per GPU-hour. Together charges SFT and DPO per million training tokens with a per-job minimum.

Can I use the OpenAI SDK with both?

Yes. Both expose OpenAI-compatible endpoints. Keep the SDK, change the base URL and model ID, then test tools, JSON, streaming and context limits. Fireworks uses api.fireworks.ai/inference/v1; Together uses api.together.xyz/v1.

Which is better for dedicated GPUs?

Together publishes lower dedicated rates on the shapes we checked: H100 $5.49 versus $8.00 and B200 $9.29 versus $13.00. Fireworks sells reserved throughput with an SLA. Compare the commitment and utilisation, not just the hourly rate.

Does either vendor guarantee capacity?

Fireworks sells reserved throughput with a throughput SLA and 99.9% uptime terms; Together sells provisioned throughput units and reserved GPU clusters. Both are quote-based.

Sources

Last checked 20 November 2026. Model versions, promotions and reserved terms change; verify the exact SKU before committing. Not paid.