Use native OpenAI and Anthropic credits with bounded accounting

AGENTS.md

@@ -50,8 +50,11 @@ infrastructure last. The agent's value is measured, not asserted.
   aggregates into `results/scorecard.md`.
 - Models (`models.py`): `openrouter/<vendor>/<model>` (OpenAI, Anthropic, Google only, per the
   technology-selection policy), `chatgpt-plan/<model>` (operator's ChatGPT plan after
-  `backintel chatgpt login`), `openai/<model>`. The judge is a different model family from
-  the analyst.
+  `backintel chatgpt login`), `openai/<model>`, and `anthropic/<model>` for native APIs.
+  Default analyst/coding worker: `openai/gpt-6-luna`; default judge: `anthropic/claude-haiku-5-5`.
+  The judge is a different model family from the analyst. Native API token usage is not a
+  per-request invoice; do not fabricate dollar charges or bypass budget admission.
+  Read `docs/native-provider-credits.md`.
 
 ## Working rules
 

README.md

@@ -29,7 +29,7 @@ not a product version. Code is available under the [MIT license](LICENSE).
 Two entry points share the same source boundaries:
 
 - **`ask`** runs the benchmark analyst and returns a structured answer with evidence, method
-  and caveats. Its default is `openrouter/openai/gpt-6-luna`.
+  and caveats. Its default is `openai/gpt-6-luna` through the native OpenAI API.
 - **`analyze`** runs an opt-in coding specialist with a persistent per-run workspace and
   artifact export. Its default is `chatgpt-plan/gpt-6.1-sol` with medium reasoning.
 
@@ -150,8 +150,12 @@ access to IntelIP's Doppler project is **not required** by the application.
 | Setting | Purpose |
 | --- | --- |
 | `BACKINTEL_ANALYSIS_MODEL` | Override the specialist default; explicit `--model` takes precedence |
-| `BACKINTEL_MODEL` | Benchmark analyst model |
-| `BACKINTEL_JUDGE_MODEL` | Method judge; default `openrouter/anthropic/claude-haiku-5.5` |
+| `BACKINTEL_MODEL` | Benchmark analyst model; default `openai/gpt-6-luna` |
+| `BACKINTEL_JUDGE_MODEL` | Method judge; default `anthropic/claude-haiku-5-5` |
+| `OPENAI_API_KEY` | Native OpenAI API key from the organization receiving API credits |
+| `ANTHROPIC_API_KEY` | Native Anthropic API key from the organization receiving API credits |
+| `OPENAI_ORG_ID`, `OPENAI_PROJECT_ID` | Optional native OpenAI account/project selectors |
+| `ANTHROPIC_WORKSPACE_ID` | Workspace selector when required by the Anthropic key |
 | `OPENROUTER_API_KEY` | Required when using the OpenRouter route |
 | `LANGSMITH_API_KEY`, `LANGSMITH_PROJECT`, `LANGSMITH_TRACING` | Optional tracing and experiment configuration |
 | `BACKINTEL_DATA_DIR` | Data directory; defaults to the checkout's `data/` |
@@ -163,6 +167,15 @@ per-request API dollar charge. There is no paid-API fallback. An OpenRouter spec
 requires both `--max-cost-usd` and `--reservation-usd`; these are application admission
 limits, not a provider-enforced billing cap. See [model selection and usage](docs/specialist-default.md).
 
+Native API calls consume the corresponding provider account's API credits. They are
+separate from the specialist's ChatGPT plan connection. Explicit model overrides still
+select OpenRouter; there is no automatic fallback. Native responses retain token usage,
+but do not supply a per-request invoice amount: `cost_usd` stays unknown. Fresh native
+improvement campaigns use separately labeled, versioned conservative token-cost
+estimates for budget admission, with incomplete usage stopping later calls. Historical
+OpenRouter accounting stays unchanged. See
+[native provider setup](docs/native-provider-credits.md) for activation and experiment boundaries.
+
 Olist acquisition additionally needs your own Kaggle access. Dataset permissions are
 separate from this repository's code license; see the [benchmark specification](docs/benchmark.md).
 

docs/native-provider-credits.md

@@ -0,0 +1,127 @@
+# Native API credits
+
+The owner requested native OpenAI and Anthropic API routing on 2026-11-13 to use
+the API credits on those accounts. This is a provider change for future runs, not
+a change to historical experiment records or permission to spend the entire balance.
+
+| Role | Default route | Credential |
+| --- | --- | --- |
+| Benchmark analyst (`ask`, benchmark harness) | `openai/gpt-6-luna` | `OPENAI_API_KEY` |
+| Optional improvement coding worker | Configured analyst model, unless explicitly overridden | Same native key by default |
+| Method judge | `anthropic/claude-haiku-5-5` | `ANTHROPIC_API_KEY` |
+| Report specialist (`analyze`) | `chatgpt-plan/gpt-6.1-sol`, medium reasoning | Existing BackIntel OAuth cache |
+
+The native routes explicitly target `api.openai.com` and `api.anthropic.com`.
+Anthropic uses the installed `langchain-anthropic` integration and native Messages
+API, including structured output. It does not pass through the OpenAI compatibility
+layer. Both routes honor completion-token, timeout and retry limits. Missing native
+credentials fail visibly; there is no fallback to OpenRouter or the plan connection.
+OpenRouter remains an explicit route for reproducing historical comparisons.
+
+## Account setup
+
+Use Doppler project `backintel`, configuration `dev`. Store the two native keys
+there, never in source, `.env`, chat or a VM mount. A browser sign-in alone does not
+configure an API key. The key must belong to the organization that receives the
+credit grant; a key from a different organization does not inherit that benefit.
+
+Optional selectors are `OPENAI_ORG_ID`, `OPENAI_PROJECT_ID`, and
+`ANTHROPIC_WORKSPACE_ID` (required for some Anthropic account-scoped keys).
+Reuse existing keys when available. Verify with each provider's read-only model
+catalog before inference, including access to the exact model IDs above.
+
+Once credentials are verified, set the existing Doppler model overrides to:
+
+```text
+BACKINTEL_MODEL=openai/gpt-6-luna
+BACKINTEL_JUDGE_MODEL=anthropic/claude-haiku-5-5
+```
+
+Explicit environment values take precedence over built-in defaults. Changing the
+code alone does not override an existing OpenRouter selection in Doppler.
+
+## Accounting and experiments
+
+Provider credits, gross API usage and cash paid are different quantities. A credit
+benefit confirmation is not proof of the current remaining balance. Native responses
+provide token counts, not the account's net dollar debit after credits.
+
+The usage tracker preserves native input/output tokens and reports unknown `cost_usd`
+with `cost_complete=false`. `NativeRunBudget` separately admits calls against a
+**conservative usage-cost estimate**, with `estimated_cost_usd`, `estimate_complete`
+and an accounting version. It never relabels that estimate as an invoice or a credit
+balance. Native budgets always report `unknown_charge=true`; continuation instead
+requires complete estimated accounting (`accounting_uncertain=false`). The original
+`RunBudget` still requires provider-reported charges for OpenRouter.
+
+The qualified policy is frozen from official standard-service prices on **2026-11-13**:
+
+| Native model | Input per million tokens | Output per million tokens |
+| --- | ---: | ---: |
+| `openai/gpt-6-luna` | $0.50 | $0.75 |
+| `anthropic/claude-haiku-5-5` | $1.00 | $2.50 |
+
+These deliberately use each model's highest standard long-context cache-write input
+rate and long-context output rate for every request. They overestimate ordinary short
+prompts and cache hits. All cached input and all output, including reasoning tokens,
+are included exactly once; discounts and credits are excluded. Each call rounds up
+to one billionth of a dollar. The policy, source URLs and model limits live in
+`src/backintel/bench/native_costs.py`; its hash is frozen into comparison plans.
+
+The native factory pins OpenAI `service_tier=default` and Anthropic
+`service_tier=standard_only` at their global endpoints. This policy covers these two
+model IDs and dated snapshots, text and local tools, up to their declared context and
+output limits. Unknown models, missing or inconsistent token counts, audio usage,
+unexpected service tiers and Anthropic paid server-tool requests cannot settle an
+estimate. An unsuccessful or unaccounted request retains its reservation and stops
+later calls. Per-call reservation overruns also stop continuation. These remain
+application admission limits, not provider-enforced billing caps.
+
+Prepare a fresh improvement campaign with native routes in both arms. Preparation
+locks the accounting basis and policy hash; execution and comparison reject changed
+or incomplete accounting. Mixed OpenRouter/native campaigns are rejected. The old
+OpenRouter dollar-complete campaign cannot be resumed by swapping keys/model IDs.
+Keep historical manifests, reported charges and unpriced-call reservations unchanged.
+
+Retain the original **$5 combined reliability allowance** across analyst, worker and
+judge calls, subtracting historical reported usage and outstanding reservations before
+setting the new run's `--max-cost-usd`. Revalidate the judge through the native adapter
+before the 92-response campaign. Available account credit does not increase that
+allowance. SDK mock tests establish adapter/accounting behavior; they do not establish
+live judge quality or actual account billing.
+
+## References
+
+- [OpenAI API credits for ChatGPT subscribers](https://developers.openai.com/api/docs/guides/chatgpt-api-credits)
+- [Anthropic Haiku 5.5 and API credit announcement](https://www.anthropic.com/claude-haiku-5-5)
+- [Native Haiku 5.5 model ID](https://platform.claude.com/docs/en/models/overview)
+- [LangChain ChatAnthropic integration](https://docs.langchain.com/oss/python/integrations/chat/anthropic)
+- [Anthropic OpenAI compatibility limitations](https://platform.claude.com/docs/en/cli-sdks-libraries/libraries/openai-sdk)
+- [OpenAI native prices](https://developers.openai.com/api/docs/pricing)
+- [Luna long-context rates and limits](https://developers.openai.com/api/docs/models/gpt-6-luna)
+- [Haiku 5.5 prices and limits](https://platform.claude.com/docs/en/models/haiku-5-5/overview)
+
+The integration reuses the already locked LangChain/Anthropic packages, with
+`langchain-anthropic` now declared as a direct dependency. This preserves the
+selected OpenAI/Anthropic model families and uses their own endpoints.
+
+## Live adapter evidence — 2026-11-13
+
+Native Anthropic completed six authored evaluator controls with complete token estimates
+and no budget uncertainty or reservation breach. Conservative usage estimate: **$0.028152**.
+The judge matched only four of six frozen criteria, so this proves the native transport
+and accounting path, not judge reliability. The larger campaign remains held pending that
+separate quality issue. [Retained LangSmith experiment](https://smith.langchain.com/o/90b24546-f3e2-4f48-af72-b06e73b27b49/projects/p/e2832945-3a35-43b1-b445-44f7b6f7c80a).
+
+Native OpenAI then completed one structured-output adapter smoke with 55 input and 13
+output tokens. Its conservative estimate was **$0.00003725**. This was a transport check,
+not an analytical question. Combined new estimates: **$0.02818925**. Actual invoice
+amounts and remaining credit balances are unknown. Evidence is retained in the canonical
+checkout under `data/output/judge-native-validation-20261113/` and
+`data/output/native-openai-smoke-20261113/`; accounting version `a0cc73ec381e4778`.
+
+The native-only working tree passed 232 tests with 15 optional skips. Integration with
+current main and the PR8 evaluator passed 247 tests with 15 optional skips, Ruff lint and
+format, Vulture and whitespace checks. The six-call trial used the same model-factory,
+usage-tracker and native-accounting bytes. Shared Doppler selectors have not been changed;
+activate native routes only after the native-capable implementation is adopted.

pyproject.toml

@@ -9,6 +9,7 @@ dependencies = [
   "deepagents==0.7.35",
   "langchain>=1.4,<2",
   "langchain-openai>=1.7,<2",
+  "langchain-anthropic>=1.7.10,<2",
   "langsmith>=0.14.17,<0.15",
   "pydantic>=2,<3",
   "polars>=2.1,<3",