Compatibility record · checked 11 August 2026

Five clients, checked against one Responses endpoint.

Infer records the exact client version, request path, bounded live result, and untested surface. A working custom base URL is not a native integration, certification, or supplier attestation.

Read the matrix Read the API contract Check live routes

pinned clients5

successful live records18 receipts · $0.026965

request path/v1/responses

current routeschecking status…

PassThe named behavior was observed for the exact version and scope in that cell.

Not testedNo client-specific evidence exists for that behavior in the current record.

Current statusLive availability refreshes separately and never rewrites a dated client result.

Client × behavior

The scope stays in every cell.

Pass never means every version or production certification. Each linked record retains its diagnostic failures, exact receipts, and limits.

Infer compatibility evidence checked 11 August 2026
Client	Pinned setup	Direct `/v1/responses`	SSE consumption	Tool round trip	Structured output	Session affinity	Live production route	Public availability
Codex CLI`0.146.0`Checked live 11 Aug 2026	PassGenerated TOML parsed locally. This proves configuration shape only.	PassCodex posted one production request to Infer’s Responses endpoint; no Chat Completions path appeared.	PassThe client consumed Infer’s Responses SSE, returned the requested text, and completed with a receipt.	Not testedNo Codex function-call and result loop was run.	Not testedNo Codex strict structured-output request was run.	Not testedThe static provider block does not add `relay.session_id`.	PassOne production-routed `infer/gpt-5.6-terra:balanced` request completed. The upstream supplier identity was not disclosed or attested.	PassThe tested selector was publicly listed and callable at the dated run. This is not an uptime promise.
OpenCode`1.18.16`Checked 11 Aug 2026	Pass`--pure debug config` resolved the custom provider. Its permissive parser makes this resolution evidence only.	PassLocal capture and the later production run both used `/v1/responses`; no Chat Completions request appeared.	PassTwo production SSE requests returned text and usage; OpenCode exited 0.	PassInfer’s real serializer completed a mocked, local read-only function-call/result loop. The live run did not exercise tools.	Not testedNo OpenCode structured-output run is in the current record.	PassTwo production requests persisted one OpenCode session through `x-session-affinity`.	PassTwo production-routed `infer/gpt-5.6-terra:balanced` requests completed. The upstream supplier identity was not disclosed or attested.	PassThe tested selector was publicly listed and callable at the dated run. This is not an uptime promise.
Pydantic AI`2.27.1`Checked live 11 Aug 2026	Pass`OpenAIResponsesModel` ran through a custom `OpenAIProvider`. Infer is not a first-class Pydantic provider.	PassAll five successful SDK calls posted to `/v1/responses`; no Chat Completions request appeared.	Pass`run_stream` consumed Infer’s buffered Responses SSE and exact text. This does not establish first-token latency.	PassOne local read-only function ran once, and the next request carried its call and result.	PassOne `NativeOutput` request used Responses-native JSON Schema and returned the typed object.	Not testedConversation state and `previous_response_id` were not exercised.	PassFive production-routed `infer/gpt-5.6-luna:low-cost` requests completed. The upstream supplier identity was not disclosed or attested.	PassThe tested selector was publicly listed and callable at the dated run. This is not an uptime promise.
LangChain OpenAI`1.4.3`Checked live 11 Aug 2026	Pass`ChatOpenAI` ran with an explicit custom base URL and `use_responses_api=True`. Infer is not a first-class LangChain provider.	PassAll six captured calls, including one uncharged diagnostic, posted to `/v1/responses`; no Chat Completions request appeared.	PassLangChain consumed Infer’s buffered Responses SSE and exact text. This does not establish first-token latency.	Pass`bind_tools` invoked one local read-only function and returned one `ToolMessage`.	PassOne native `json_schema` request returned a validated Pydantic object.	Not testedServer conversation state and `previous_response_id` were not exercised.	PassFive production-routed `infer/gpt-5.6-terra:balanced` requests completed. The upstream supplier identity was not disclosed or attested.	PassThe tested selector was publicly listed and callable at the dated run. This is not an uptime promise.
Vercel AI SDK`7.0.59@ai-sdk/openai 4.0.37`	Pass`createOpenAI` ran with a custom Infer base URL and explicit `.responses()`. This is not a native Infer provider.	PassEvery captured AI SDK call posted to `/v1/responses`; no Chat Completions request appeared.	Pass`streamText` consumed Infer’s buffered Responses SSE and exact text. This does not establish first-token latency.	PassOne read-only local function executed once, followed by one tool-result turn.	PassOne strict `Output.object` request mapped a Zod schema to Responses-native JSON Schema.	Not tested`previousResponseId` and Conversations were not exercised.	PassFive production-routed `infer/gpt-5.6-terra:balanced` requests completed. The upstream supplier identity was not disclosed or attested.	PassThe tested selector was publicly listed and callable at the dated run. This is not an uptime promise.

Pass covers only the named successful run. Paid and uncharged diagnostic failures remain on the linked record pages; they are not hidden by a passing cell.

Evidence ledger

Four records. Eighteen receipts.

The successful records total $0.026965 in customer charges. Diagnostic attempts and their charges remain itemized on the linked pages.

01 · coding agents

Codex + OpenCode

Three streamed production requests produced three receipts: Codex one, OpenCode two in one persisted session. Successful charge total: $0.018303.

Read all three receipts

02 · Python SDK

Pydantic AI 2.27.1

Five completed calls covered text, buffered SSE, one local read-only tool loop, and one native schema. Successful charge total: $0.000071.

Read the Pydantic record

03 · Python framework

LangChain OpenAI 1.4.3

Five completed calls covered text, buffered SSE, one local read-only tool loop, and one native schema. Successful charge total: $0.004369.

Read the LangChain record

04 · TypeScript SDK

Vercel AI SDK 7.0.59

Five completed calls covered generateText, streamText, one local read-only tool loop, and Output.object. Successful charge total: $0.004222.

Read the AI SDK record

Evidence boundary

What the records do not prove.

Every pass is pinned to one client version, date, selector, and bounded workload.

Native integrationThe SDK/framework paths use their OpenAI Responses clients with a custom Infer base URL. They are not first-class Infer providers or endorsements.

Streaming latencyThe clients consumed valid Responses SSE. Infer currently buffers the upstream response, so these checks do not establish first-token latency.

Tool breadthThe passing tool rows cover one local read-only function. Hosted tools, state-changing tools, parallel calls, approvals, MCP, and arbitrary schemas remain untested.

State and recoveryAsync paths, retries, cancellation, failover, long runs, images, server conversation state, and future versions remain outside this record.

Supplier identityInfer receipts establish its routed request record and customer charge. They do not disclose or attest the upstream supplier or model weights.

Exact client recordsEach page retains its setup, receipts, diagnostics, and untested boundaries.

Codex →OpenCode →Pydantic AI →LangChain →Vercel AI SDK →

Start with one capped request.

Create an account, verify the email, add service credit, and check the selected route before running a pinned setup.

Create account Read the Responses contract