OpenAI Fine-Tuning vs Together AI vs Fireworks AI: Real Fine-Tuning Cost per 100 Million Training Tokens in 2026
Before you spend hours comparing token price cards across fine-tuning providers, there is an uncomfortable reality you need to verify first: whether the vendor will actually let you train your model today. OpenAI has closed self-serve fine-tuning to new customers, leaving only eligible existing accounts to run jobs during an active deprecation wind-down. Meanwhile, alternative platforms like Together AI and Fireworks AI publish transparent token rate cards, but finding a cheap per-token training rate does not guarantee that your target model architecture, training method, or serving infrastructure is available on both platforms.
This guide separates account access gates, billable training token volumes, validation multipliers, and ongoing inference deployment costs. All figures below reflect published platform documentation checked on 4 November 2026, calculated to reveal real project budgets rather than synthetic marketing claims.
OpenAI: account eligibility and model lifetime are separate checks
OpenAI’s official availability timeline restricts new fine-tuning exclusively to previously active customers. Since 2 July, organizations that have not actively consumed a fine-tuned model for inference within the preceding 60 days are completely blocked from creating new fine-tuning jobs. Even for eligible accounts, the final cutoff to create new training jobs arrives on 6 January 2027, unless your chosen base model snapshot retires earlier.
Crucially, do not confuse the training creation cutoff with the model serving deadline. The listed GPT-4o, GPT-4o mini, GPT-4.1, and GPT-4.1 mini snapshots—along with all custom models fine-tuned on top of them—shut down permanently on 14 January 2027. Specific smaller snapshots, including GPT-4.1 nano and o4-mini checkpoints, already shut down on 23 October 2026. If you fine-tune an OpenAI model today, you are investing engineering hours into an artifact with a strictly bounded operational lifespan.
The pricing page still displays an o4-mini training row at $100 per hour. However, that listing does not override its official shutdown notice or establish a purchasable fine-tuning option today. The SFT guide separately confirms the existing-customer restriction window. We do not convert an hourly rate to an estimated token price or invent speculative future pricing for retired endpoints.
The buying decision: If you are a new customer, do not build an architecture around self-serve OpenAI fine-tuning—the door is closed. If you are an eligible existing customer, require a confirmed job quote and a documented migration plan before sinking budget into a model snapshot that will shut down in January 2027. Neither scenario justifies claiming that OpenAI fine-tuning has completely vanished overnight, but neither supports building new long-term production systems on it.
Published training rates: use the model and method, not a vendor average
All figures below reflect USD per one million billable tokens. Together AI’s rows represent Llama 3.1 8B from its current model-specific rate table. Fireworks AI’s rows show its published managed-training tier for models up to 16B parameters. A broad parameter-tier rate is never confirmation of an identical base checkpoint, dataset formatting support, or equivalent downstream task accuracy. Always confirm exact model checkpoint availability before treating a rate card as a binding quote.
| Published rate card | SFT LoRA | SFT full | DPO LoRA | DPO full |
|---|---|---|---|---|
| Together: Llama 3.1 8B | $0.34 | $0.38 | $0.84 | $0.94 |
| Fireworks: up to 16B managed-training tier | $0.50 | $1.00 | $1.00 | $2.00 |
Sources: Together pricing, LoRA and full fine-tuning tabs and Fireworks managed training pricing. Together enforces a $4 minimum charge per training job for this model. Older size-tier pricing cards should not be substituted for these current model-specific rates. Furthermore, notice that full parameter fine-tuning on Together costs only a modest premium ($0.38 vs $0.34) over LoRA for Llama 3.1 8B, whereas Fireworks doubles the rate ($1.00 vs $0.50). There is no universal 2.5x multiplier across vendors.
Fireworks also maintains separate Serverless Training API and Dedicated Training API products. Their token-based or GPU-hour billing units must not be conflated with this managed SFT and DPO table. Promotional discounts, startup credits, and free tiers are excluded from these calculations.
What does 100 million training tokens mean?
In fine-tuning arithmetic, confusion frequently arises around the phrase "100 million tokens." The critical formula is processed training tokens = dataset tokens × epochs. If 100 million represents the total processed tokens across the entire run, you do not multiply by epochs again. If 100 million describes your raw dataset size, running a standard 3-epoch training schedule processes 300 million tokens before factoring in validation passes or evaluation steps.
| Training-only workload | Together LoRA | Together full SFT | Fireworks LoRA SFT tier | Fireworks full SFT tier |
|---|---|---|---|---|
| 100 million processed tokens | $34.00 | $38.00 | $50.00 | $100.00 |
| 100 million dataset tokens × 3 epochs | $102.00 | $114.00 | $150.00 | $300.00 |
These numbers isolate raw training compute charges. They deliberately exclude data curation, synthetic data generation, validation passes, hyperparameter retries, local storage, and model deployment fees. A lower training compute fee never guarantees that one model will converge to acceptable benchmark scores at the same dataset size or epoch count as another.
Validation passes and job minimums alter small-job budgets
Together AI’s billing documentation explicitly states that billable tokens include training tokens across all epochs plus validation tokens across all evaluation steps. The exact token count is determined by the platform’s tokenizer at upload. For the selected LoRA configuration, your planning formula is max($4, [(dataset tokens × epochs) + (validation tokens × evaluations)] ÷ 1,000,000 × $0.34).
Consider a practical scenario: a 10-million-token dataset trained over 3 epochs, paired with a 1-million-token validation set evaluated 4 times throughout the run. This generates 34 million billable tokens: 30 million training tokens plus 4 million validation tokens. At Together’s LoRA rate ($0.34/M), the actual invoice is $11.56, rather than the $10.20 you would budget for training tokens alone. Conversely, a tiny experimental run of 100,000 dataset tokens across 3 epochs without validation produces only 300,000 billable tokens ($0.102 on paper), meaning the $4.00 job minimum controls your checkout payment.
On Fireworks, intermediate reasoning-trace workloads and chain-of-thought fine-tuning require careful token counting. Multi-turn reasoning examples expand token counts dramatically compared to standard single-turn supervised datasets. Never apply a generic multiplier across dissimilar prompt formats; verify token counts against your actual JSONL dataset files.
The training invoice is only the down payment
A fine-tuned model that sits on an idle dedicated server without handling inference requests burns cash rapidly. Together’s documentation warns that dedicated inference endpoints continue billing for every hour they remain active, whether they process requests or sit completely idle. Fireworks advertises fine-tuned model serving at base-model token rates, but that applies only when serverless deployment is supported for that specific model architecture and adapter size. You must obtain clear quotes for deployment mode, cold-start latency, dedicated capacity minimums, and idle hourly costs before committing.
Always budget fine-tuning as an end-to-end engineering initiative: dataset preparation and cleansing + multiple training and evaluation runs + inference hosting over the product lifetime + eventual model migration. For instance, a dedicated GPU endpoint billed at a modest $0.50 per hour left running for a 30-day billing month accumulates $360.00 in idle hosting fees. That $360 hosting overhead dwarfs a $15 or $20 difference on your initial training invoice.
Earlier versions of industry guides frequently used disputed historical OpenAI rates to claim massive cost savings by switching platforms. We reject those sweeping generalizations: comparing different base models, unaligned training objectives, and restricted-access endpoints cannot establish a like-for-like financial saving. Base your provider decision on empirical task accuracy, serving latency, and total lifecycle costs.
A four-step trial to make the right buying decision
- Define strict task benchmarks. Maintain a held-out test dataset with an objective scoring rubric tailored to your domain. Benchmark the baseline off-the-shelf model before investing in fine-tuning, and measure the exact delta in accuracy.
- Document the exact technical parameters. Record the base model snapshot, parameter count, SFT or DPO objective, LoRA rank or full fine-tuning configuration, tokenizer specifics, epoch count, evaluation schedule, and quoted rate card. Note any proprietary lock-ins.
- Budget every iteration and failure. In machine learning, your first run is rarely your final production model. Plan for hyperparameter sweeps, bad data splits, and rejected candidate checkpoints. The true cost of fine-tuning is the cost per accepted production model, not the cost of a single script run.
- Test deployment throughput and exit strategies. Deploy the candidate adapter to a staging endpoint and measure real-world inference latency, concurrency limits, and idle billing rules. Verify that you can export model weights cleanly without vendor lock-in.
The bottom line: For any new fine-tuning initiative, shortlist open-weight models with proven serving pathways, run small measured pilot jobs, and budget for the complete deployment lifecycle. Existing OpenAI customers should audit their remaining snapshot deadlines and map out a migration strategy before the January 2027 cutoff. For infrastructure alternatives, explore our guide on self-hosted fine-tuning costs and our comprehensive pricing methodology.