LLM Inference Cost Comparison 2026: DigitalOcean vs. Together AI, Fireworks AI, Modal, Nebius, Baseten, and OpenRouter

DigitalOcean charges the lowest modular input value among providers we comparison present for real-time gpt-oss-120b conclusion astatine $0.10/million tokens. Together AI and Fireworks AI some complaint $0.15/million input tokens for the aforesaid model. They complaint a little output value of $0.60/million tokens compared pinch DigitalOcean’s $0.70 per cardinal output tokens.

Fireworks AI has the lowest value for asynchronous processing for the straight comparable gpt-oss-120b script modeled beneath owed to the discount connected inputs and outputs provided by their batch API. The main competitor group for this article is DigitalOcean, Together AI, Fireworks AI, Modal, Nebius, Baseten, and OpenRouter. OpenAI is only shown for pricing discourse for their ain GPT models and is not considered a like-for-like inference-platform competitor.

This LLM inference costs comparison is based connected publically disposable and verified prices arsenic of July 30, 2026. Prices are ever changing, truthful please reappraisal against charismatic pricing pages anterior to publication and refresh quarterly. This article compares the 7 conclusion platforms connected their respective pricing models: per-token serverless APIs, routed exemplary access, managed endpoints, and compute-based deployments. Because platforms whitethorn connection different models, deployment modes, and billing structures, this article flags non-comparable rows alternatively of forcing them into a azygous user ranking based connected unverifiable normalization assumptions.

Which supplier is cheapest for what?

The lowest-cost conclusion supplier depends connected the model, token distribution, latency requirements, batch eligibility, and measured cache-hit rate. The pursuing array compares 5 accumulation scenarios utilizing publically listed prices and intelligibly defined workload assumptions.

Prices are listed successful US$ / 1 cardinal tokens. Estimates do not see taxes, grounded aliases retried requests, web fees, vector retention costs, aliases negotiated discounts. Prices and exemplary readiness are taxable to alteration by the provider. Verified linked pricing pages for up-to-date prices.

DigitalOcean is competitory for real-time open-model conclusion aliases immoderate exertion looking to colocate inference, storage, databases, and retrieval infrastructure. Fireworks is well-suited for batch jobs and repeated prefixes. Together offers an expansive exemplary catalog on pinch serverless and dedicated deployment paths.

Modal and Nebius matter if you want to really deploy and optimize a exemplary onto metered GPU infrastructure alternatively than buying a token-priced API. Baseten offers some per-token Model APIs and paths for dedicated and self-hosted deployment. OpenRouter is an aggregation/routing furniture whose full costs depends connected the selected model/provider value positive its platform-fee rules. OpenAI provides nonstop entree to proprietary GPT models and is retained beneath only arsenic a pricing context.

How the afloat inference-platform competitor group compares

The array beneath therefore, compares purchasing models and directs readers to each provider’s unrecorded pricing page.

OpenAI and Anthropic are bully conceptual value anchors for proprietary models; however, they are contextual model-lab comparisons alternatively than members of the superior inference-platform competitor group utilized here.

How overmuch does LLM conclusion costs per cardinal tokens?

The schematic beneath breaks down really text-generation API pricing is calculated based connected tokens consumed arsenic input/output. Additional calculations are shown based connected caching tokens and utilizing different work features.

Keep this look successful mind erstwhile shopping astir for LLM conclusion providers, but make judge to reappraisal each provider’s up-to-date pricing for caching, tooling/function calls, web searches, storage, location usage, and dedicated infrastructure.

Consider gpt-oss-120b.The cheapest LLM API depends connected the equilibrium betwixt input and output tokens. Using 10,000 input tokens and 100 output tokens, DigitalOcean would costs astir $0.00107, which is astir 31% cheaper than Together AI / Fireworks astatine $0.00156. Using 100 input tokens and 10,000 output tokens, Together AI / Fireworks would costs astir $0.00 6015, which is astir 14% cheaper than DigitalOcean astatine $0.00701.

DigitalOcean has a little costs for input-intensive workloads, and Together AI and Fireworks person a little costs for output-intensive workloads. The providers scope the break-even constituent erstwhile the input-token measurement is doubly the output-token volume.

Pricing table: Cost per 1 cardinal tokens successful July 2026

The array provides a comparison of emblematic serverless input/output prices for DigitalOcean, Together AI, Fireworks AI, and Baseten, straight listing the aforesaid model. Modal and Nebius are included successful the array supra for platforms, arsenic their deployment economics alteration based connected metered compute utilization; OpenRouter is included arsenic an illustration routing layer, arsenic its effective value is based connected the selected model/provider way and level interest rules. OpenAI is shown only successful a proprietary exemplary pricing context. Prices were verified connected July 30, 2026, and whitethorn alteration complete time.

Important pricing qualifications

• The 3 numbers successful Fireworks’ pricing bespeak uncached input / cached input/output.
• For GPT-5.6 Luna, Terra, and Sol, the displayed prices use to prompts of nary much than 272,000 tokens.
• DigitalOcean states that its OpenAI (commercial-model) integration uses customers’ OpenAI API keys, pinch billing done by OpenAI.
• Claude Sonnet 5’s $2/M input and $10/M output prices are introductory. On September 1, 2026, Anthropic’s published modular prices go $3/M input and $15/M output.

This comparison uses straight published serverless database prices and excludes batch discounts, privilege tiers, negotiated contracts, taxes, storage, web charges, grounded requests, and dedicated GPU deployments. Verify the linked pricing pages earlier making purchasing aliases architecture decisions.

Note that this array is not attempting to beryllium a value ranking. Two providers whitethorn beryllium serving checkpoints pinch different quantization, discourse windows, throughput, aliases reliability. Identical exemplary names from the aforesaid supplier do not guarantee operational equivalence. Record complete exemplary identifiers and benchmark pinch the aforesaid prompts, sampling parameters, output magnitude caps, concurrency, and information dataset.

What is the cheapest LLM API for batch processing millions of documents?

Batch conclusion processes files aliases groups of requests asynchronously. Results do not request to beryllium returned immediately, truthful providers tin schedule activity much optimally and complaint less. Batch useful good for classification, metadata extraction, moderation, evaluations, offline summarization, synthetic-data generation, embedding pipelines – fundamentally thing that’s not interactive chat aliases latency-sensitive agents.

DigitalOcean’s Inference pricing documentation explicitly states “up to 50%” for supported OpenAI and Anthropic models. Their Batch Inference guide specifications the workflow, the 24-hour expected completion, and explicitly states the commercialized models’ supported scope. The Serverless Inference overview plainly states token-based serverless billing and erstwhile you’d want to usage that complete dedicated inference. Fireworks AI states that batch input/output is priced astatine 50% of serverless pricing, making it cheaper than a provider’s real-time price. OpenAI’s Batch API likewise offers 50% little costs than synchronous API processing for supported models and endpoints.

When you’re evaluating operationally, comparison much than the percentage. Test record size limits, occupation quotas, completion timeouts, cancellation, partial failures, ordering of results, idempotency, retries, observability, etc.

Scenario 1: Classify 1 cardinal documents overnight

The pursuing little illustration demonstrates really you mightiness usage the prices from the comparison array supra pinch an existent workload. For instructions and guidance connected erstwhile and really to usage batch vs real-time inference, spot DigitalOcean’s Introducing Unified Batch Inference and The LLM Inference Trilemma: Throughput, Latency, Cost. These articles explicate batch implementation and the trade-offs progressive successful choosing an conclusion endpoint, while this page provides existent pricing comparisons and workload calculations.

Suppose a institution must categorize 1 cardinal support tickets earlier the adjacent business day. The exemplary returns a category, urgency score, and short explanation.

Assumptions:

• 1,000,000 documents
• 800 input tokens and 30 output tokens per document
• 800 cardinal input tokens and 30 cardinal output tokens total
• gpt-oss-120b
• No prompt-cache discount
• Fireworks uses batch rates
• DigitalOcean and Together usage modular serverless rates because the DigitalOcean open-model batch discount is not assumed.

The pursuing illustration demonstrates the costs savings of classifying 1 cardinal documents pinch gpt-oss-120b nether modular serverless rates and discounted batch conclusion pricing. Fireworks Batch reduces the full value to $69, fixed the workload assumptions below.

If you’re a interrogator performing conclusion complete ample datasets, Fireworks Batch offers the lowest estimated costs for this example. This intends that batch conclusion tin person important costs savings for investigation usage cases that aren’t time-sensitive (such arsenic classifying datasets, generating synthetic data, exemplary evaluation, aliases large-scale archive analysis). However, researchers should ever measure token value alongside exemplary versions, output quality, reproducibility, time-to-completion, nonaccomplishment recovery, complaint limiting, and more.

Scenario 2: Run a chatbot astatine 10,000 requests per day

Let’s activity done different illustration utilizing a customer-support chatbot that needs to watercourse responses. We can’t usage batch conclusion successful this lawsuit because customers request responses immediately.

Assumptions:

• 10,000 requests per time * 30 days = 300,000 requests per month
• 1,200 input tokens per petition * 300,000 requests per period = 360 cardinal input tokens per month
• 300 output tokens per petition * 300,000 requests per period = 90 cardinal output tokens per month
• gpt-oss-120b

Without punctual caching

A chatbot accepting 10,000 requests/day utilizing gpt-oss-120b will costs $99/month connected DigitalOcean and $108/month connected Together aliases Fireworks earlier caching, tools, retries, and the infrastructure astir that.

How punctual caching changes the winner

Chatbot inputs see repeated strategy instructions, personification information policy, instrumentality definitions, output schema, illustration interactions, and merchandise context. Assume 70% of the 360 cardinal monthly input tokens suffice for Fireworks’ published $0.015 cached-input rate:

• 108 cardinal uncached input tokens
• 252 cardinal cached input tokens
• 90 cardinal output tokens

Prompt caching tin importantly trim your LLM conclusion by reusing prompts pinch communal contented specified arsenic strategy instructions, information policies, instrumentality definitions, and output schema definitions. Below, we show really a 70% cache-hit complaint reduces a chatbot’s estimated monthly costs from $108 to $73.98.

This is not intended to beryllium an apples-to-apples cache comparison; It reveals sensitivity to a published discount. You must verify each provider’s cache support for the circumstantial model, minimum prefix length, lifetime, routing behavior, cache constitute cost, and whether prefixes tin beryllium shared betwixt users. Note the resulting cache-hit ratio. A timestamp that often updates, user-specific data, aliases a dynamically generated database of devices adjacent the commencement of the punctual tin inhibit reuse.

Scenario 3: Embeddings and RAG complete 100,000 documents

A retrieval-augmented generation strategy has 3 main costs:

1. Converting documents into embeddings.
2. Storing and searching those embeddings successful a vector database.
3. Sending the retrieved accusation to an LLM to make answers.

Consider you person 100,000 documents pinch 1,000 tokens each. This equals 100 cardinal tokens. At $0.009 per cardinal tokens for all- MiniLM-L6-v2, the first embedding costs would be: 100 cardinal tokens×$0.009=$0.90

Creating embeddings for each 100k documents only costs $0.90! This doesn’t mean the full RAG strategy will costs $0.90. Additional expenses whitethorn see Document storage, vector database hosting, query embeddings, reranking, LLM input/output tokens, and archive re-indexing.

Let’s dress your exertion receives 1 cardinal questions per month, pinch each mobility averaging 20 tokens. The query embeddings will embed 20 cardinal tokens per month: 20×$0.009=$0.18

Most of the costs will astir apt travel from utilizing the LLM. If the RAG strategy successfully retrieves 2,000 tokens for each mobility the personification asks, past those cardinal questions will nonstop 2 cardinal retrieved tokens to the model. At $.10 / cardinal tokens for input, the retrieved discourse will cost: 2,000×$0.10=$200

This $200 only applies to the retrieved discourse sent to the model. The value of the user’s questions and the tokens the exemplary generates are not included.

Retrieval settings tin frankincense effect the last bill. The much documents you retrieve, and the larger and much overlapping your chunks are, the much tokens you nonstop to the LLM. Metadata filters and reranking tin trim irrelevant output earlier procreation occurs. Evaluate retrieval accuracy, reply quality, and costs together.

Verdict: Generating embeddings for 100,000 documents pinch 100 cardinal tokens only costs $0.90 here. The LLM procreation and vector-database infrastructure will apt costs acold much successful a accumulation RAG strategy than this first upfront costs for archive embeddings.

Cheapest measurement to tally DeepSeek aliases Llama 70B successful production

For Llama 3.3 70B, DigitalOcean charges $0.65 per cardinal input tokens and $0.65 per cardinal output tokens. Together AI charges $1.04 per cardinal tokens for some input and output. Let’s ideate our chatbot uses 360 cardinal input tokens and generates 90 cardinal output tokens per month. Our monthly full measurement would be: 360+90=450 cardinal tokens

Because each supplier bills input and output tokens astatine the aforesaid rate, our monthly value for each value tin beryllium easy calculated:

DigitalOcean: 450×$0.65=$292.50

Together AI: 450×$1.04=$468

DigitalOcean would costs $175.50 little per period for this workload: $468−$292.50=$175.50. This translates to savings of 37.5% compared pinch Together AI’s $468/month value tag. However, this calculation only compares token prices. It does not bespeak if some providers meet the aforesaid latency, throughput, reliability, aliases value of generated text. To decently comparison vendors, the aforesaid exemplary version, prompts, procreation settings, and workload should beryllium tally connected both.

DeepSeek could mention to the distilled type trained connected apical of Llama, the ample mixture-of-experts model, the Flash version, aliases the Pro version. For instance, present are the prices DigitalOcean lists for the DeepSeek model:

• DeepSeek R1 Distill Llama 70B: $0.99 input and $0.99 output per cardinal tokens.
• DeepSeek V4 Flash: $0.112 input and $0.224 output.
• DeepSeek V4 Pro: $1.392 input and $2.784 output.

Note that these models each person different capabilities, architectures, and prices and truthful should not beryllium clustered into a generic “DeepSeek” label. Remember to log the afloat exemplary identifier + type erstwhile comparing providers.

The exemplary pinch the lowest costs per token whitethorn not consequence successful the lowest costs per completed task. A cheaper but little tin exemplary whitethorn nutrient incorrect answers, require longer prompts, return aggregate retries, aliases request quality intervention. Benchmarking by value per token is important, but users should besides see full costs to get an acceptable consequence erstwhile comparing providers. Infographic showing really to measurement existent costs per successful AI task by accounting for input, output, retry, and instrumentality costs:

For sustained usage cases, dedicated conclusion could outperform token billing. DigitalOcean is advertizing an AMD MI300X astatine $2.59/hour and an NVIDIA H100 astatine $4.41/hour.

However, a GPU-hour value doesn’t show you overmuch without measured throughput and utilization. A adjuvant break-even calculation is to disagreement the hourly costs of an endpoint by measuring successful tasks per hr to comparison against the serverless costs per successful task.

What the per-token value does not show you

Price unsocial shouldn’t find your prime of conclusion provider. Consider really the pursuing aspects of their operations mightiness impact your accumulation costs and capacity much heavily:

• Rate limits and capacity: Low prices aren’t valuable if your exertion can’t get capable requests aliases tokens per minute. Look into defaults, approved quotas, burst behavior, concurrency limits, and whether higher quotas require committed usage. Microsoft Azure AI Together provides dynamic, per-model limits that standard pinch sustained traffic. Batch queues whitethorn person different record and occupation quotas.
• Latency and acold starts: Measure clip to first token for interactive applications, including inter-token latency, end-to-end latency, P95 and P99, timeouts, and acold starts. Good mean latency whitethorn obscure a damaging tail. For example, a one-second hold whitethorn beryllium catastrophic successful a chat interface but irrelevant for overnight batch jobs.
• Reliability and quality: Track completed transactions alternatively than HTTP 200 responses. Count rate-limit retrials, invalid JSON, incorrectly formed instrumentality calls, quiet completions, information blocks, hallucinations, and failures astatine your exertion layer. Compare suppliers utilizing the aforesaid information set. Test many times astatine concurrency levels you expect to spot successful production.
• Model and serving differences: The aforesaid family sanction tin disguise different versions, quantization formats, discourse windows, kernels, and decoding defaults. Pin afloat identifiers erstwhile possible. Providers whitethorn update aliases aliases discontinue checkpoints, truthful support regression tests and a controlled upgrade procedure.

Practical strategies for reducing LLM conclusion costs

Practical costs optimization spans pricing decisions, workload design, exemplary routing, retrieval quality, measurement, and operational discipline. Here is simply a applicable model to thief trim expenses without compromising reliability aliases value of answers.

Conclusion

There is nary cheapest LLM API crossed each accumulation workloads successful 2026. As of value verification connected July 29th, DigitalOcean has the lowest listed modular input complaint for gpt-oss-120b among providers, compared to a important listed-price advantage for Llama 3.3 70B. That illustration chatbot costs $99/month connected DigitalOcean vs. $108/mo connected Together aliases Fireworks anterior to caching effects. For chatbot measurement connected Llama 3.3 70B, DigitalOcean is $292.50 compared to Together’s $468.

Fireworks offers unbeatable worth erstwhile your eligible workloads tin usage its batch aliases cached-input pricing tiers. The 1 million-document illustration costs drops from $138 astatine modular Fireworks rates to $69 pinch Batch Pricing, and assuming a 70% cache-hit rate, reduces the chatbot estimate to $73.98. Together remains competitory successful catalog and deployment flexibility, and it comes retired up connected selected models for illustration Qwen 3.7 Plus. Baseten fits successful this nonstop Model API comparison, wherever it lists the aforesaid model. Modal and Nebius request utilization-aware compute benchmarking. OpenRouter requires route- and fee-aware costing. OpenAI is being utilized arsenic a proprietary exemplary value anchor, not arsenic a superior inference-platform competitor successful this comparison.

Once you person measurable constraints typical of production, the durable purchasing rule is simple: optimize for costs per successful task alternatively than costs per cardinal tokens successful isolation. Measure exemplary version, input/output mix, cache-hit ratio, latency distribution, retries, quality, surrounding infrastructure, and tally a mini production-shaped benchmark. Choose the supplier (or routing strategy) that gives you the lowest reliable value nether existent exertion constraints.

FAQ

What is the cheapest LLM API successful 2026?

There is nary universally cheapest API. In this comparison, DigitalOcean is the cheapest for the stated real-time gpt-oss-120b and Llama 3.3 70B workloads; Fireworks is the cheapest for the eligible gpt-oss-120b batch and high-cache scenarios; Together wins selected catalog rows. Baseten is straight comparable connected overlapping Model APIs; Modal and Nebius must beryllium evaluated from measured compute utilization; OpenRouter must beryllium evaluated utilizing its selected way and applicable fee; and OpenAI is due erstwhile proprietary GPT capabilities are required.

How should I comparison LLM API prices?

Multiply input and output volumes by their respective rates, adhd cached tokens, retries, tools, and infrastructure, past disagreement by successful tasks. Use identical prompts, exemplary versions, output limits, concurrency, and value thresholds crossed providers.

Can batch conclusion trim LLM costs by 50%?

Yes, for eligible models and workloads. Fireworks publishes a 50% batch discount connected serverless input and output. OpenAI publishes abstracted batch rates astatine half modular rates for supported models. DigitalOcean advertises up to 50% connected supported OpenAI and Anthropic models, not each unfastened model.

When is dedicated conclusion cheaper than serverless?

Dedicated capacity tin beryllium cheaper nether sustained, predictable request and precocious utilization. Benchmark successful tasks per GPU-hour connected the target hardware, including idle clip and operations, and comparison that worth pinch the serverless costs per successful task.

References

• Inference Pricing
• Scalable, Cost-Efficient AI: Introducing Unified Batch Inference connected DigitalOcean
• Together AI Pricing In 2026: Models, Costs, And How To Manage Your Bill