Providers

Native batch availability is separate from model, region, and account eligibility. A provider can expose an API while rejecting a particular model. The current adapters use the following control planes; deterministic fixtures cover their wire lifecycles, while authenticated qualification remains provider-specific.

Connection / codec	Input	Cancel	List	Partial results	Qualification note
OpenAI Chat and Responses	JSONL file	Yes	Yes	No	Chat submit/results/cancel qualified live; Responses codec has deterministic coverage.
Anthropic Messages	Inline request set	Yes	Yes	No	Live job completed with two keyed successes; item errors, cancellation, and expiry remain distinct.
Mistral Chat	File or inline	Yes	Yes	No	Both modes returned keyed live outcomes; inline output can mirror the output file.
Gemini native	File or inline	Yes	Yes	No	Both modes returned keyed live outcomes; operation API uses BATCH_STATE_*.
Qwen / Model Studio	JSONL file	Yes	Yes	No	Singapore qwen-turbo returned keyed live outcomes; region/model eligibility still matters.
xAI / Grok Chat	Sealed JSONL file	Yes	Yes	Yes	Sealed file path returned keyed live outcomes; failed-item envelope remains unqualified.
Groq Chat	JSONL file	Yes	Yes	No	The configured account returned not_available_for_plan at file upload.
Together Chat	JSONL file	Yes	Yes	No	Live job accepted and still validating; listing has no documented continuation cursor.
Fireworks control plane	JSONL dataset	No	Yes	No	Explicitly composed adapter supports submit/status/list; typed results remain unsupported pending a real output-record sample.
Bedrock Claude 3 Haiku	S3 JSONL file	Yes	Yes	No	Optional AWS SDK transport, service role and buckets; 100-record minimum.

Fireworks uses an explicit FireworksBatchDriver with LLMConfig, FireworksBatchSettings(accountId: ...), BatchHttpTransport, and CanUploadBatchFile. Wrap it in BatchRuntime to use the common facade. The driver creates and uploads an account-scoped dataset before creating the job; the job request names a dedicated output dataset. Its saved reference can resume status reads in another process. A failed job creation reports the input dataset, proposed output dataset, and job resource names. The inspected public documentation does not define its complete output-record envelope, so results() raises UnsupportedBatchOperation before HTTP and job snapshots report BatchResultsAvailability::Unsupported. The documented DELETE operation is not treated as cancellation. The Fireworks key available for qualification created and uploaded a dataset, but job creation returned a payment-method-required error. The direct DeepSeek API has no native batch contract in the inspected public API; hosted DeepSeek models may be batched through an eligible host such as Model Studio.

capabilities()->canReadResults() is false for Fireworks until its result record codec is qualified; it is true for drivers that decode typed item outcomes.

Choose provider options explicitly where they affect the native job:

<?php
use Cognesy\Polyglot\BatchInference\Config\GeminiBatchOptions;
use Cognesy\Polyglot\BatchInference\Config\MistralBatchOptions;
use Cognesy\Polyglot\BatchInference\Config\QwenBatchOptions;
use Cognesy\Polyglot\BatchInference\Enums\GeminiBatchInputMode;
use Cognesy\Polyglot\BatchInference\Enums\MistralBatchInputMode;

$mistralOptions = new MistralBatchOptions(MistralBatchInputMode::Inline, timeoutHours: 24);
$geminiOptions = new GeminiBatchOptions(GeminiBatchInputMode::File, displayName: 'nightly');
$qwenOptions = new QwenBatchOptions(completionWindow: '48h', taskName: 'nightly');

$job = $batches->submit($items, $mistralOptions); // use a Mistral connection
// @doctest id="d6e3"

OpenAI uses OpenAIBatchOptions, Groq uses GroqBatchOptions, and xAI uses XAIBatchOptions. Options from a different provider are rejected before submission. capabilities()->inputModes() reports the adapter's enforced item/byte/record limits for each supported form. These limits do not certify that a model is batch-enabled.

Bedrock uses its own AWS control plane and S3 rather than an LLM API-key preset. Install aws/aws-sdk-php in the consuming application, configure AWS credentials through its normal SDK provider chain, and provision a Bedrock service role plus input and output buckets. Compose it explicitly:

<?php
use Cognesy\Polyglot\BatchInference\BatchInference;
use Cognesy\Polyglot\BatchInference\BatchRuntime;
use Cognesy\Polyglot\BatchInference\Config\BedrockBatchOptions;
use Cognesy\Polyglot\BatchInference\Config\BedrockBatchSettings;
use Cognesy\Polyglot\BatchInference\Drivers\Bedrock\BedrockBatchDriver;
use Cognesy\Polyglot\BatchInference\Transport\AwsSdk\BedrockBatchSdkTransport;

$settings = new BedrockBatchSettings(
    region: 'us-east-1',
    roleArn: 'arn:aws:iam::123456789012:role/BedrockBatchRole',
    inputBucket: 'my-batch-input',
    outputBucket: 'my-batch-output',
);
$driver = new BedrockBatchDriver($settings, BedrockBatchSdkTransport::forRegion($settings->region()));
$batches = BatchInference::fromRuntime(new BatchRuntime($driver));
$job = $batches->submit($items, new BedrockBatchOptions(timeoutHours: 48));
// @doctest id="dc8a"

The current Bedrock codec supports only the Claude 3 Haiku InvokeModel Messages body. It rejects tools, structured output and reasoning requests. The 100-record minimum and 100,000-record adapter cap are surfaced through capabilities()->inputModes(). AWS account, region and model quotas can impose additional limits. Output records preserve recordId, model output or error, and S3 provenance. Bedrock SDK requests are covered by deterministic mocks; this adapter has not been run against an authenticated AWS account. retrieve() reports result availability as pending until results() checks the job-specific S3 output prefix; completion alone does not prove retained objects exist.

When Mistral returns outputs inline alongside an output_file, the adapter uses the inline rows and does not replay their file duplicates. It can still read a separate error file and skips any keyed row already delivered inline.

For Qwen, the reference binds the selected Beijing or Singapore control-plane URL. The Singapore qwen preset's current default model is not in the documented Singapore batch list; select a listed batch model or an eligible Beijing connection. Qwen also requires homogeneous models and thinking mode within an input file. The specific API reference permits a 6 MB JSONL line, while its summary page says 1 MB; live qualification is still needed for the disputed range.

Source contracts: OpenAI, Anthropic, Mistral, Gemini, Qwen, xAI, Groq, Together, and Fireworks, and Bedrock.