Maximum capability
GPT‑6 Astra and Claude Fable 5.1 are for the hardest long-running work. Use them when a wrong answer or failed agent run costs more than the tokens.
Checked 23 September 2026
AI models compared: price, strengths and best use
The best model is the least expensive one that passes your own evaluation. This guide maps the current model families to real workloads and shows when the premium tiers are worth their price.
The short version
Best default tier
GPT‑6 Sol, Claude Sonnet 5 and Gemini 3.8 Flash offer strong quality at much lower prices. Start here for code, analysis and agentic work.
High volume
GPT‑6 Luna, Claude Haiku 4.5 and Gemini Flash‑Lite suit classification, extraction, moderation and simple worker tasks at scale.
These are API prices, not monthly ChatGPT, Claude or Gemini app subscriptions. Prices are USD per one million text tokens at the standard service tier. A subscription buys access to a product with usage limits; an API bills actual consumption. Caching, batch processing, tools, long context and faster service tiers can change the bill.
Current general-purpose models
| Provider and model | Input / output | Context | Best for | Verdict |
|---|---|---|---|---|
| OpenAI GPT‑6 AstraFlagship | $10 / $50 | 1.05M | The hardest end-to-end tasks, demanding analysis and agent runs | Highest capability, but 5× Sol's price. Use selectively. |
| OpenAI GPT‑6 SolCoding and agents | $2 / $10 | 1.05M | Complex coding, tool use and professional knowledge work | Our default OpenAI choice. |
| OpenAI GPT‑6 LunaEfficiency | $0.10 / $0.50 | 1.05M | Focused, well-specified tasks at high volume | Exceptionally cheap; escalate hard cases. |
| Claude Fable 5.1Maximum Claude capability | $10 / $50 | 1M | Long-running research, difficult reasoning and autonomous agents | Use when Opus at high effort still fails the eval. |
| Claude Opus 5.5Agentic work | $4 / $20 | 1M | Long-running coding, refactoring, computer use and knowledge work | Anthropic's recommended start for most heavy workloads. |
| Claude Sonnet 5Balanced | $2 / $10 | 1M | Everyday code, analysis, content, vision and tool use | Best Claude price/performance for routine professional work. |
| Claude Haiku 4.5Fastest | $1 / $5 | 200K | Low latency, high volume and lower-cost subagents | Start here when speed matters more than maximum depth. |
| Gemini 3.8 FlashCurrent stable model | $0.75 / $3.75Introductory price through 2026 | 1.05M | Multimodal analysis, long coding tasks, agents and Google tools | Aggressively priced; rises to $1.50 / $7.50 in 2027. |
| Gemini 3.1 Pro PreviewPro preview | $1 / $5Introductory price through 2026 | 1.05M | Advanced multimodal understanding and agent prototyping | Not our first production choice while it remains preview. |
| Gemini 3.5 Flash‑LiteHigh volume | $0.30 / $2.50 | 1.05M | Translation, simple data processing and many small agent tasks | Low cost, though Luna is cheaper per text token. |
Scope: this table covers the main general text and multimodal choices for new projects. It excludes retiring legacy models and specialist image, video, speech, embedding and moderation models. “M” means million tokens. Thinking or reasoning tokens are normally billed as output.
How to choose without overspending
Start in the middle tier
Run GPT‑6 Sol, Claude Sonnet 5 or Gemini 3.8 Flash on 20–50 real tasks with identical criteria and blind scoring.
Move down before moving up
Run Luna, Haiku or Flash‑Lite on the same set. If the cheaper model passes just as often, the difference is pure savings.
Escalate only difficult cases
Use rules or an inexpensive router model to send only hard work to Astra, Fable or Opus. This often beats using one premium model for everything.
Measure cost per accepted result
Token prices hide retries, unnecessarily long answers and extra tool calls. Divide the complete bill by accepted tasks, not raw tokens.
What the family names mean
OpenAI: Astra is maximum capability, Sol is the workhorse for coding and agents, and Luna is the efficient small model. All three publish the same large context window and core tool support, so the choice is mainly difficulty, latency and price. For one million input plus 250,000 output tokens, list cost is about $22.5 on Astra, $4.5 on Sol and $0.225 on Luna.
Anthropic: Fable sits above the traditional Opus tier for the longest and hardest runs. Opus is the heavy workhorse, Sonnet balances intelligence and price, and Haiku prioritises latency. Claude is especially relevant when long documents, rich prose quality or long-running agent work dominate the workload.
Google: Flash no longer means a small model. Gemini 3.8 Flash is Google's current all-round choice for code, agents and multimodal input. Flash‑Lite is the volume tier. Pro Preview is interesting for experiments, but a preview endpoint is less predictable than a stable generally available model.
Do not choose from one public benchmark. Vendors use different test setups, and results change with tools, effort, language and prompt. A small representative internal evaluation is worth more than a generic leaderboard.
Reasoning controls differ
All three platforms have moved reasoning control out of prompt phrasing and into a parameter. OpenAI exposes reasoning effort, Anthropic an effort setting, and Google a thinking level. Opus 5.5 keeps adaptive thinking on and defaults to medium effort. Start with the documented default, test one level lower, and increase it only where the acceptance rate improves enough to cover the extra output tokens and latency.
Using two model families together
OpenAI models
Prompting Astra, Sol and Luna, choosing reasoning effort, caching and model escalation.
Read the guide →Claude models
Fable, Opus, Sonnet and Haiku, effort, long context and structured output.
Read the guide →Gemini models
Gemini 3.8 Flash, Pro Preview and Flash‑Lite, thinking levels and production status.
Read the guide →GPT‑6 + Claude Code
A practical workflow for letting ChatGPT, Codex and Claude Code plan, build and review the same project without file collisions.
Read the guide →Primary sources
- OpenAI model comparison — capability, context and pricing
- OpenAI API pricing — standard, cache, batch and long-context rates
- Anthropic models overview — current family, context and recommended starting point
- Anthropic pricing — tokens, caching and tools
- Google Gemini models — status and model IDs
- Google Gemini API pricing — current and announced rates