The short answer
For financial analysis, pick the tier before the vendor. A 10-K-sized job (100K tokens in, 4K out) costs from $0.012 on GPT-6 Luna to $1.20 on Claude Fable 5.1; the median premium model costs about 52x the median budget model.
For financial analysis, the choice among Claude, GPT and Gemini is a tier decision before it is a vendor decision. On a 10-K-sized job (100K tokens in, 4K out), the current lineups run from $0.012 per filing on GPT-6 Luna to $1.20 on Claude Fable 5.1; the median premium model costs about 52x the median budget model.
The current lineup, priced per filing
| Model | Vendor | Input $/Mtok | Output $/Mtok | Cached input $/Mtok | Context | ~$ per 10-K (100K in + 4K out) |
|---|---|---|---|---|---|---|
| Claude Fable 5.1 | Anthropic | $10 | $50 | $0.25 | 1M | $1.20 |
| GPT-6 Astra | OpenAI | $10 | $50 | $1 | 1.1M | $1.20 |
| GPT-5.5 | OpenAI | $5 | $30 | $0.50 | 1.1M | $0.62 |
| Claude Opus 5.5 | Anthropic | $4 | $20 | $0.20 | 1M | $0.48 |
| Gemini 3.1 Pro Preview | $2 | $12 | $0.20 | 1M | $0.25 | |
| Claude Sonnet 5.5 | Anthropic | $2 | $10 | $0.10 | 1M | $0.24 |
| GPT-6.1 Sol | OpenAI | $2 | $10 | $0.10 | 1.1M | $0.24 |
| Gemini 4 Argon | $2 | $10 | $0.10 | 1M | $0.24 | |
| GPT-5.4 Mini | OpenAI | $0.75 | $4.50 | $0.075 | 400K | $0.093 |
| Gemini 3.8 Flash | $0.75 | $3.75 | $0.075 | 1M | $0.090 | |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | $0.03 | 1M | $0.040 | |
| Claude Haiku 5.5 | Anthropic | $0.10 | $0.50 | $0.01 | 1M | $0.012 |
| GPT-6 Luna | OpenAI | $0.10 | $0.50 | $0.01 | 1.1M | $0.012 |
Per-filing costs are list price × tokens, computed from the table, not a benchmark run. Prices are the vendors' standard (non-batch) list rates as recorded in the aidevhub.io model catalog, whose rows were verified against each vendor's pricing page (catalog prices verified 2026-10-09). Some rates carry conditions: GPT-6 Astra: prompts up to 272K input tokens; GPT-5.5: prompts up to 272K input tokens; Gemini 3.1 Pro Preview: prompts up to 200K tokens; GPT-6.1 Sol: prompts up to 272K input tokens; Gemini 4 Argon: promotional price through November 30, 2026; Gemini 3.8 Flash: promotional price through December 31, 2026; Claude Haiku 5.5: prompts up to 100K tokens; 5-minute cache writes; GPT-6 Luna: prompts up to 272K input tokens.
What the prices tell you
Pick the tier first. Premium: Claude Fable 5.1 ($1.20), GPT-6 Astra ($1.20), GPT-5.5 ($0.62), Claude Opus 5.5 ($0.48), Gemini 3.1 Pro Preview ($0.25), Gemini 4 Argon ($0.24). Mid: Claude Sonnet 5.5 ($0.24), GPT-6.1 Sol ($0.24), GPT-5.4 Mini ($0.093), Gemini 3.8 Flash ($0.090). Budget: Gemini 3.5 Flash-Lite ($0.040), Claude Haiku 5.5 ($0.012), GPT-6 Luna ($0.012). The median premium model costs about 52x the median budget model, so the tier a task needs decides most of its cost.
Then compare vendors like for like. Each vendor's cheapest option per tier: premium, Anthropic Claude Opus 5.5 $0.48, OpenAI GPT-5.5 $0.62, Google Gemini 4 Argon $0.24; mid, Anthropic Claude Sonnet 5.5 $0.24, OpenAI GPT-5.4 Mini $0.093, Google Gemini 3.8 Flash $0.090; budget, Anthropic Claude Haiku 5.5 $0.012, OpenAI GPT-6 Luna $0.012, Google Gemini 3.5 Flash-Lite $0.040. Where two vendors price a tier the same, the choice is capability on your documents, not price.
Output is where long answers cost money. Every model in the table charges at least 5x more per output token than per input token. A 10-K extraction is input-heavy, so the input rate decides it; a long research memo is output-heavy, so check the output column before the input column.
Read the conditions before you budget. A promotional rate expires on a stated date, and rates that apply to prompts up to a length step up above it for the whole request. Both are in the table's source rows; the cost tools on this site apply them.
Where each tier fits in finance work
- High-volume field extraction (every numeric field from thousands of filings): the budget tier. At Gemini 3.5 Flash-Lite ($0.040), Claude Haiku 5.5 ($0.012), GPT-6 Luna ($0.012), a market-wide sweep is affordable, and a reasoning tier only needs to see the subset that fails validation.
- Mid-tier reasoning at scale (ratio analysis, comparing several filings): the mid tier. Claude Sonnet 5.5 ($0.24), GPT-6.1 Sol ($0.24), GPT-5.4 Mini ($0.093), Gemini 3.8 Flash ($0.090).
- Full-filing synthesis and hard judgment (a merged filing set, a long memo, an investment thesis): the premium tier. Claude Fable 5.1 ($1.20), GPT-6 Astra ($1.20), GPT-5.5 ($0.62), Claude Opus 5.5 ($0.48), Gemini 3.1 Pro Preview ($0.25), Gemini 4 Argon ($0.24).
- A filter in front of a frontier model: run the cheap tier over everything and pass only the flagged documents up. The per-filing gap above is the saving.
Caching changes the math for repeated context
If every call resends the same large block (a system prompt, a filing-structure schema, a peer-filing brief), cached input decides the bill. Cached input rates on the current lineup range from $0.01 to $1 per million tokens (column 5). How to Set Up Prompt Caching for Finance works the break-even math, and the Token Cost Optimizer prices your own prompt mix with and without the cache.
Verified engine output
The block below runs the Model Selector for Finance on the hardest finance profile: full-filing synthesis, sub-30-second latency, top budget. The selector reads the same catalog as the table above, so its pick and its budget estimate move with it.
Decision guidance
- Classify the task tier: extraction, mid reasoning, or frontier reasoning. This is the dominant cost lever.
- Pick the cheapest model in that tier that clears your accuracy bar on your own documents.
- Run a small eval to break vendor ties inside a tier, where list prices are often identical.
- Add caching if a large fixed context repeats across calls, and check a promotional rate's end date before you commit a budget to it.
Connects to
- Model Selector for Finance: the tier-fit engine behind the recommendation.
- Best LLM for Financial Analysis 2026: the task-tiered pillar.
Sources
Model prices are the vendors' list prices, read from the pages below and kept in the model price catalog the cost tools share.
- Anthropic, platform.claude.com/docs/en/about-claude/models/overview: Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 (read 2026-10-16).
- Anthropic, platform.claude.com/docs/en/about-claude/pricing: Claude Haiku 5.5 (read 2026-10-16).
- Google, ai.google.dev/gemini-api/docs/pricing: Gemini 3.1 Pro Preview, Gemini 4 Argon, Gemini 3.8 Flash, Gemini 3.5 Flash-Lite (read 2026-10-18).
- OpenAI, developers.openai.com/api/docs/pricing: GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna (read 2026-10-16).
- OpenAI, developers.openai.com/api/docs/models/gpt-5.5: GPT-5.5 (read 2026-10-16).
- OpenAI, developers.openai.com/api/docs/models/gpt-5.4-mini: GPT-5.4 Mini (read 2026-10-16).
Verified engine output
Show the recompute-verified inputs and outputs
| task | synthesize |
|---|---|
| latency | sub_30s |
| cost | b200_plus |
| context | k200_1m |
| quality | high |
| ranked (13 items) | [...] |
|---|
Computed live at build time.
Frequently asked questions
- Which is cheapest for financial analysis: Claude, GPT or Gemini?
- On list price, GPT-6 Luna (OpenAI) at $0.10 input and $0.50 output per million tokens, about $0.012 for a 100K-token 10-K. The budget tier as a whole: Gemini 3.5 Flash-Lite ($0.040), Claude Haiku 5.5 ($0.012), GPT-6 Luna ($0.012).
- Is Claude or GPT better value at the top tier?
- Premium-tier costs per 10-K: Claude Fable 5.1 ($1.20), GPT-6 Astra ($1.20), GPT-5.5 ($0.62), Claude Opus 5.5 ($0.48), Gemini 3.1 Pro Preview ($0.25), Gemini 4 Argon ($0.24). Each vendor's cheapest premium option: Anthropic Claude Opus 5.5 $0.48, OpenAI GPT-5.5 $0.62, Google Gemini 4 Argon $0.24. Where the prices match, break the tie with an evaluation on your own documents.
- Do they all fit a full 10-K in context?
- Yes. Every model in the current lineup has at least a 400K-token context window, and a 10-K is roughly 100K tokens. Some rates step up above a prompt length; the table's conditions say where.
- How much does caching save?
- Cached input is billed at a fraction of the standard input rate on every vendor in the table (column 5 gives each model's rate). For a large fixed preamble resent on every call, the cached path wins after a few reuses.
- Where do these prices come from?
- From the aidevhub.io model catalog, which records each model's list price with the vendor URL it was verified on (catalog prices verified 2026-10-09). Per-filing costs are computed from those prices, not from a benchmark run.