AI models comparison 2026: text, code, image, video, voice

An AI models comparison is only useful if it compares like with like and says when it was read. This page puts the main models of each modality — frontier text, coding, small and fast, image, video and voice — in one table each, with the vendor's own list rate, context window and limits as published on 19 October 2026. Each table links to the deeper comparison for that tier, and the last section is the part no table can do for you: how to compare AI models on your own workload.

Every rate below is the model vendor's list rate: what an account without a negotiated agreement is billed by that vendor. If your organization has negotiated rates, those replace it. For the full catalog of models with reference rates, searchable and sortable, see the model directory.

How to read these tables

Three rules keep a comparison table honest.

• Units differ by modality. Text is billed per million tokens, most image models per image or per megapixel, video per second of output, and speech per character or per hour. Do not set a per-token number beside a per-image number.
• Rates move on schedules. Several rates below change by prompt length, time of day or date. The tables give the base rate and name the condition; the deeper comparisons work through the arithmetic.
• "Not stated" means not stated. Where the vendor page read on 19 October did not show a fact, the table says so rather than carrying an older figure forward.

Frontier text and reasoning models

These are the models each vendor positions as its most capable for reasoning and long, complex work.

Two lines in that table matter more than the headline rate. GPT-6 Astra and Claude Fable 5.1 list the same $10 and $50, but OpenAI reprices a request above 272K input tokens while Anthropic bills the whole 1M window at the standard rate¹³. And Gemini 3.1 Pro is still published under a preview model code⁶, a lifecycle fact that belongs in any production decision. The frontier models comparison covers caching, batch and regional uplifts for the top four.

Coding models

Coding agents resend large prompts many times per task, so context, caching and prompt-length tiers shape the bill more than in chat.

One change since the coding models comparison was written: on 19 October Alibaba Cloud's international pricing page no longer listed qwen3-coder-plus; the only Qwen3-Coder model on it was qwen3-coder-flash¹¹. If you built a cost model on the coder-plus tiers, check the model is still offered in your region before relying on it.

Small and fast models

The tier that carries most production traffic by volume: classification, extraction, routing and short summaries.

Claude Haiku 5.5 is not in the small and fast models comparison. Its rate depends on prompt length: a request over 100,000 tokens pays the higher rate on the whole request, cache reads and writes included³. DeepSeek's peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays⁹.

Open-weight models

Open weights separate the lab that trained a model from the host that runs it, so the rate depends on who hosts it. The facts that travel with the model are its license and its limits.

For the Chinese labs' families, where each publishes weights, and the hosting and compliance questions, see the open-weight models guide.

Image generation models

Image models bill in three different units, so the column that matters is the unit.

Retirements belong in this table too. OpenAI shuts down gpt-image-1 on 23 October 2026 and gpt-image-1.5 and gpt-image-1-mini on 1 December 2026, naming the GPT Image 2.5 models as replacements¹⁸. The image models comparison covers sizes, reference images and how results come back.

Video generation models

Video is billed per second of output, and resolution usually moves the rate.

The Veo 3.1 line is the reason a video comparison needs a date: Google's deprecation page lists 22 October 2026, three days after this page was read, as the shutdown date for the three Veo 3.1 preview model ids, and notes that listed dates are the earliest possible ones¹⁹. Kling, Seedance and Runway are covered in the video models comparison; their rates were not re-read for this page.

Voice: speech and transcription models

Speech is the modality where the billing unit varies most within one vendor: OpenAI prices tts-1 per character and gpt-4o-mini-tts per token¹. Convert to cost per minute of finished audio at your own average speaking rate before comparing.

How to compare AI models on your own workload

The tables tell you which models can run a task and what the provider will bill per unit. They cannot tell you which model does the task well enough, or what a successful result costs. That needs your own data.

1. Filter on hard constraints first. Context window, output limit, input modalities, region, lifecycle status (preview, retiring) and license. This usually removes half of any table above.
2. Build an evaluation set from real traffic. Fifty to a few hundred real inputs per task type, each with a check a machine or a reviewer can apply: schema valid, tests pass, label matches, image accepted. The guide to evaluating LLMs on real workloads covers how to build one.
3. Run every shortlisted model on the same inputs, under the same conditions. Same prompt, same tools, same effort level, the stable prefix cached the way you would ship it.
4. Record estimated provider cost and outcome per request. Include retries and failed attempts; they are billed too.
5. Compute cost per successful task. Total estimated cost divided by the number of tasks that passed. The companion piece on cost per successful task goes deeper.
6. Re-run on a dated trigger. A rate change, a new model or a retirement on the calendar is a reason to repeat the run, not to trust last quarter's result.

Comparing models through INFRO

INFRO does not score outputs; your evaluation harness does. What it gives a comparison is one OpenAI-compatible endpoint in front of the provider accounts you connect, so every candidate is called the same way, and a record of each request with its estimated provider cost. Your providers bill your accounts directly at your rates; INFRO's subscription is separate and nothing it sells is priced per token.

• Run candidates side by side. The console Playground runs text, image, video, speech and transcription on one of your connections or through a gateway key, and copies the call as cURL.
• Estimate from your rates. Each response carries usage.cost, an estimate of what your provider will bill, and price_source. Enter negotiated rates under Prices (billing) so the comparison ranks rates you actually pay, not list rates.
• Group the run. GET /v1/usage?group_by=model returns requests, estimated cost, unpriced requests, error rate and latency percentiles per model (usage analytics).
• Ship the decision as configuration. A route names the chosen model first and a tested alternative second, so a retirement like Veo 3.1's is a console edit, not a code change (routing).

The catalog carries one rate per model, so a request above a vendor's long-prompt threshold or in DeepSeek's off-peak window is estimated at the base rate; check those against the provider's own usage data.