# AI for tradies: model data and editorial provenance

Checked 11 October 2026. This is a frozen metadata snapshot for this editorial guide. There were no authenticated requests, paid model calls, trade-task tests or model-generated test outputs.

## Snapshot method

Read the unauthenticated [OpenRouter model catalogue](https://openrouter.ai/api/v1/models) and select the eight exact IDs below. `inputPrice` and `outputPrice` are `pricing.prompt` and `pricing.completion` multiplied by 1,000,000. `context` is `context_length`. All selected records had a null `expiration_date` at retrieval; this is not a promise of future availability.

The [OpenRouter models documentation](https://openrouter.ai/docs/guides/overview/models) defines prices as USD per token/request/unit, with top-provider values and possible conditional overrides. Cards show uncached base text rates. They omit cache pricing, request/tool/media charges, reasoning-token volume, route changes and taxes. They are not prices for ChatGPT, Claude or Gemini app subscriptions.

| OpenRouter ID | Canonical slug in snapshot | Input USD / 1M | Output USD / 1M | Context tokens |
|---|---|---:|---:|---:|
| `openai/gpt-6-astra` | `openai/gpt-6-astra-20260903` | 10.00 | 50.00 | 1,050,000 |
| `anthropic/claude-opus-5.5` | `anthropic/claude-opus-5.5-20260921` | 4.00 | 20.00 | 1,000,000 |
| `openai/gpt-6.1-sol` | `openai/gpt-6.1-sol-20260929` | 2.00 | 10.00 | 1,050,000 |
| `anthropic/claude-sonnet-5.5` | `anthropic/claude-sonnet-5.5-20260928` | 2.00 | 10.00 | 1,000,000 |
| `google/gemini-3.8-flash` | `google/gemini-3.8-flash-20260902` | 0.75 | 3.75 | 1,048,576 |
| `deepseek/deepseek-v4.1-flash` | `deepseek/deepseek-v4.1-flash-20260910` | 0.30 | 1.20 | 1,048,576 |
| `anthropic/claude-haiku-5.5` | `anthropic/claude-haiku-5.5-20261007` | 0.10 | 0.50 | 1,000,000 |
| `openai/gpt-6-luna` | `openai/gpt-6-luna-20260922` | 0.10 | 0.50 | 1,050,000 |

Canonical slugs are OpenRouter identifiers, not independently established provider release dates or necessarily request-pinnable versions. Card IDs preserve the catalogue IDs. In a future evaluation, separately record the exact served provider/version.

## Official provider cross-checks

- [OpenAI: GPT-6 Astra](https://developers.openai.com/api/docs/models/gpt-6-astra) confirms the name, demanding-work positioning, $10/$50 base rates and 1,050,000-token context. For prompts over 272,000 input tokens, full-request input rates double and output rates rise by 50%.
- [OpenAI: GPT-6.1 Sol](https://developers.openai.com/api/docs/models/gpt-6.1-sol) confirms the name, complex professional-work positioning, $2/$10 base rates and 1,050,000-token context. The OpenRouter snapshot includes the same long-prompt multiplier as Astra.
- [OpenAI: GPT-6 Luna](https://developers.openai.com/api/docs/models/gpt-6-luna) confirms the name, focused/high-volume positioning, $0.10/$0.50 base rates and 1,050,000-token context. The same long-prompt multiplier applies above 272,000 input tokens.
- [Anthropic: Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5) confirms the identity, complex-work positioning and $4/$20 base rates. Anthropic's published comparative performance claims are not reproduced as Web Blend scores.
- [Anthropic: Claude Sonnet 5.5](https://www.anthropic.com/claude-sonnet-5-5) confirms the identity, everyday scoped-work/document positioning and $2/$10 base rates. Its original launch page lists older cache-read prices; the later Haiku announcement below updates Sonnet cache reads. Cache rates are not displayed on this guide.
- [Anthropic: Claude Haiku 5.5](https://www.anthropic.com/claude-haiku-5-5) confirms the identity and narrow/repeated-task positioning. Its rates are $0.10/$0.50 for prompts up to 100,000 tokens and $0.50/$2.50 above that threshold. This page also reduces Sonnet 5.5 cache reads to $0.10 per million.
- [Google: Gemini 3.8 Flash](https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash) confirms the stable model ID, text/image/video/audio/PDF inputs, 1,048,576 input-token limit and structured output. Displayed prices are from OpenRouter rather than inferred from this page.
- [DeepSeek: Models and pricing](https://api-docs.deepseek.com/quick_start/pricing/) confirms `deepseek-flash` currently serves DeepSeek V4.1 Flash, with vision, JSON output and tool calls. Direct prices depend on the time/day: $0.15/$0.60 off-peak and $0.30/$1.20 peak for uncached input/output. The guide deliberately displays the separate OpenRouter snapshot value of $0.30/$1.20; it must not claim this is always the cheapest direct rate.

Input modalities and structured-output support in card copy were also checked against `architecture.input_modalities` and `supported_parameters` in the selected OpenRouter records. These describe API routes; they do not guarantee identical app capabilities.

## Editorial choices and limits

Tier placement and task matching are Web Blend editorial hypotheses, not findings from tests. S+/S are frontier-oriented classes, A/B are workhorse classes and C is a lightweight class. The S-versus-A and A-versus-B distinctions are deliberately coarse editorial placement; evidence here cannot establish a strict cross-provider performance order. There is no score, accuracy percentage, winner badge, tie-breaker or claim to have established global state of the art.

S+ contains Astra and Opus because their providers position them for the most involved work. S contains Sol as a broad complex-work option. A contains Sonnet and Gemini as everyday/workflow candidates. B contains DeepSeek as a lower-cost repeatable-work candidate. C contains Haiku and Luna because their providers explicitly position them for focused, repeated work. Price alone does not establish capability, and a C candidate may be the most useful option for the selected task.

`bestFor` means “include in a first trial for this task”, not “tested best”. Task filters should preserve the same tiers and metadata date. Trade filters should change only examples and constraints. No field in this file can justify trade-specific winners, public-upload capability, autonomous sending, ledger posting or compliance decisions.

The job examples are invented illustrations with no customer data, measured outcomes or real-job endorsements. All five task types require a suitable human reviewer. The recommended next validation is a small owner-usability exercise, then a separately scoped, recorded task trial.

Provider icon sources and attribution are recorded in /ai-guide/providers/provenance.md. OpenRouter popularity rankings, dataset reuse and third-party benchmark licences are outside this snapshot; no popularity data is included.