GPT, Claude or Gemini? Comparing Five AI Model Families and Their Costs
A dated comparison of GPT, Claude, Gemini, DeepSeek and Qwen: inputs, tools, API costs and limits, with practical criteria for writing, coding and analysis.
Starting with “Which model ranks first?” is an easy way to choose badly. Writing, changing code, analyzing video and processing a hundred thousand documents need different abilities. A model solving a few more difficult questions doesn’t necessarily justify a higher price for every summary.
My recommendation: choose by task, then test with your own material. Start routine work with a moderately priced model and reserve flagships for problems the cheaper option cannot handle.
The original article checked official catalogs and pricing pages available on October 7, 2026, covering OpenAI, Anthropic, Google, DeepSeek and Alibaba Cloud Qwen. Names, capabilities and prices below retain that snapshot; translation is not a new verification. Recommendations interpret those sources, not cross-vendor hands-on testing or an independent ranking of vendor claims.
GPT: a complete toolchain, but cost the flagship carefully
The cited OpenAI catalog positions GPT-6 Astra for complex reasoning and coding, GPT-6.1 Sol for capability-cost balance, and GPT-6 Luna for cost-sensitive frequent tasks. It lists text and image input, text output and a 1.05M-token context window for all three.
The attraction is the toolchain: search, file retrieval, function calls and computer use. For workflows joining research, file reading, actions and output, those interfaces reduce integration work.
Flagship cost is a limitation. Standard short-context output is listed at $50 per million tokens for Astra and $10 for Sol. Astra may suit difficult reasoning, but summaries, tagging and first drafts needn’t all use it. Tools have separate costs and configuration; support for search doesn’t mean every request automatically verifies facts.
Image input in a text model doesn’t mean the same interface generates images, video or real-time speech. Those use separate specialized models.
Try first for: tool-enabled assistants and mixed code/document tasks. Evaluate Sol, then check whether Astra actually reduces rework.
Claude: positioned for long tasks, but unit price isn’t the whole bill
The source lists Claude Fable 5.1, Opus 5.5 and Sonnet 5.5, respectively positioned for demanding reasoning and long-running agents, sustained coding and knowledge work, and speed-capability balance, each with 1M context.
Project changes, organizing multiple documents and preserving requirements across turns fit that positioning better than a single impressive answer. For repeated editing, assess whether revisions preserve meaning and constraints.
It does not establish that Claude is always the best writer. Test tone and language on your own drafts. A model may complicate simple prose or quietly alter facts.
Listed base input/output prices are $4/$20 per million tokens for Opus and $10/$50 for Fable. Their always-on adaptive thinking may add time and usage. Anthropic also notes that newer tokenizers may produce more tokens for identical text, so identical token rates don’t guarantee identical bills.
Try first for: cross-file changes, long documents and sustained editing. Start with Sonnet 5.5 if speed and cost matter more.
Gemini: broad input support; distinguish understanding from generation
The cited stable Gemini 3.8 Flash supports text, images, video, audio and PDFs as input, with text output. It lists 1,048,576 input tokens and 65,536 output tokens, plus search grounding, file search, code execution and structured output.
That combination suits mixed material, such as a vehicle launch video, product PDF and demonstration screenshots. Extract sourced points, then have an editor check them. Explicit input support is the advantage; it doesn’t prove the highest accuracy on every video question.
The model doesn’t directly generate images or audio and doesn’t support Live API. Choose specialized models for real-time speech or media generation. Video interpretation still depends on token budgets, files and sampling; a large context window doesn’t guarantee every detail is recognized.
Gemini 3.1 Pro is marked Preview in the cited catalog. Production users need to check release status and retirement rules; Latest aliases may switch versions.
Try first for: combined video, audio and PDF analysis, or workflows needing Google search grounding.
DeepSeek: low API prices, with model-specific capabilities
The source maps deepseek-flash to DeepSeek-V4.1-Flash and deepseek-v4-pro to DeepSeek-V4-Pro-0813. Both list 1M context, thinking and non-thinking modes, tool calls and JSON output.
Low prices make batch classification, summaries, extraction and preliminary screening worth testing. If quality meets requirements, savings can fund flagship review for harder cases.
Cheap calls don’t replace acceptance checks. Repeated retries or extensive manual corrections can still make a task expensive. Peak and off-peak pricing also differ; the lowest price isn’t an all-day rate.
One important distinction in the cited documentation: Flash supports vision; Pro does not. “Pro” doesn’t imply a superset of Flash’s capabilities. Compatible API styles also don’t preserve every vendor-specific feature.
Try first for: high-volume text and structured tasks. Verify visual support for image work.
Qwen: a natural candidate for existing Alibaba Cloud users
The cited Model Studio recommendations include qwen3.8-max, qwen3.7-plus and qwen3.8-flash, alongside image/video understanding and Omni models.
For teams already using Alibaba Cloud accounts, permissions and billing, the centralized integration path is practical. Trial Chinese business documents, content classification and knowledge questions before adding another provider.
Different models and service forms have different limits. qwen3.8-omni-flash handles audio/video analysis and text generation; real-time conversation uses corresponding realtime models. Don’t generalize one model’s features across the family. Prices, regions and quotas depend on the actual service, so the original article omits Qwen prices that weren’t verified on the same basis.
Cloud APIs and downloadable weights are separate. Open-source families don’t mean the cloud flagship can be privately deployed; inspect weights, licenses and hardware requirements. Third-party DeepSeek, GLM and Kimi models in Model Studio aren’t Qwen models.
Try first for: Chinese applications on existing Alibaba Cloud infrastructure and teams centralizing model service management.
API prices: a shortlist aid, not a final bill
Figures are US dollars per million tokens, using direct API uncached input and output. GPT uses Standard short-context pricing, Claude base rates and Gemini Standard paid rates. Subscription, tools, taxes, caching and batch discounts are excluded.
| Model | Input | Output | Conditions |
|---|---|---|---|
| GPT-6 Astra | 10 | 50 | Standard, short context |
| GPT-6.1 Sol | 2 | 10 | Standard, short context |
| GPT-6 Luna | 0.10 | 0.50 | Standard, short context |
| Claude Opus 5.5 | 4 | 20 | Base rate |
| Claude Sonnet 5.5 | 2 | 10 | Base rate |
| Gemini 3.8 Flash | 0.75 | 3.75 | Listed rate through December 31, 2026 |
| DeepSeek-V4.1-Flash | 0.15 / 0.30 | 0.60 / 1.20 | Off-peak / peak, uncached input |
| DeepSeek-V4-Pro-0813 | 0.66 / 1.32 | 1.98 / 3.96 | Off-peak / peak, uncached input |
OpenAI’s long-context tier costs more; don’t use these rates for a full-window request. Google lists $1.50/$7.50 for Gemini 3.8 Flash starting January 1, 2027. DeepSeek peak hours are Monday–Friday outside Chinese statutory holidays, UTC 01:00–04:00 and 06:00–10:00; other hours are off-peak.
Measure the cost of a completed task: input, visible output, billable thinking tokens, tools, retries and manual revisions. Tokenizers count the same Chinese text differently, so character count isn’t enough.
Apply the comparison to real work
| Task | Evaluate first | Check |
|---|---|---|
| Coding and tool use | GPT-6.1 Sol, Claude Sonnet 5.5; upgrade for hard tasks | Tests, unintended changes, total time |
| Editing and long documents | Claude, GPT, optionally Qwen | Preserved facts, constraints, revision effort |
| Launch-event video and audio | Gemini 3.8 Flash, relevant Qwen Omni | Timestamps, sources, figures and names |
| Batch summaries and extraction | DeepSeek Flash, GPT Luna, Qwen Flash | Acceptance rate, format stability, total cost per item |
AI and driver-assistance reporting needs careful checks of dates, versions, regions and conditions. A feature available on one vehicle in one city doesn’t imply availability across every model worldwide. Brand reputation cannot replace source checking.
Prepare ten real documents and give candidates identical requirements. Check supported conclusions, copied figures, required rewrites, elapsed time and cost to reach an acceptable result. That is more useful than asking models to rate one another.
My order: verify input types and service conditions, compare acceptance rates, then calculate total cost. Readers must understand writing, code must run, and reporting must be verifiable. A brand isn’t an acceptance criterion.
Official sources
These are the original article’s October 7, 2026 sources; models and prices may change.
- OpenAI: models, pricing.
- Anthropic: model overview, pricing.
- Google: models, Gemini 3.8 Flash, pricing.
- DeepSeek: models and pricing.
- Alibaba Cloud: Model Studio models.