TheRouter.aiTheRouter.ai
Back to Models

Grok 4.20

xaixai/grok-4.20-non-reasoning

xAI's flagship model (March 2026) with 2M token context window. Fast general-purpose mode without reasoning overhead, supporting text and image inputs with function calling and structured outputs.

xAI models overview and pricing β†—xAI release notes β†—

Grok 4.20 Non-Reasoning is the low-latency variant of xAI's Grok 4.20 Beta generation, released via the xAI API in March 2026. It uses the same underlying model weights as the reasoning variant but deliberately skips the extended chain-of-thought (CoT) step, trading deep multi-step reasoning for significantly faster time-to-first-token and higher throughput. The model retains a 2-million-token context window β€” the largest among the Grok 4.20 family β€” and accepts both text and image input.

From a TheRouter routing perspective, xai/grok-4.20-non-reasoning maps to the pinned API slug grok-4.20-0309-non-reasoning. The Grok 4.20 generation sits between the original Grok 4 (July 2025, 256K context) and the current flagship Grok 4.3 (April 2026, 1M context, always-on reasoning). As of June 2026, the Grok 4.20 variants remain active and listed on the xAI API, though xAI now recommends Grok 4.3 for new chat and coding integrations. The non-reasoning mode is the right choice when response latency matters more than reasoning depth β€” automated pipelines, high-volume batch processing, or tool-calling loops where the model is one step in a larger workflow.

The Grok 4.20 Beta generation launched as four concurrent API models: the default grok-4.20-beta (balanced), grok-4.20-0309-non-reasoning (speed), grok-4.20-0309-reasoning (depth), and grok-4.20-multi-agent-0309 (reliability via four parallel agents). All four share the same $2/$6 per 1M token pricing at launch and the same 2M-token context window. Community benchmarks (Artificial Analysis, June 2026) show the non-reasoning variant scores significantly lower than the reasoning variant on intelligence benchmarks (AA Intelligence Index 29.7 vs the reasoning variant's higher score) but delivers approximately 200 tokens per second β€” consistent with the reasoning variant's raw throughput once you discount reasoning wait time.

Best for
  • β€’ High-throughput agentic pipelines β€” reasoning is skipped, so the model acts as a fast function-calling node; ideal when the orchestrator handles logic and the model handles tool dispatch and text generation
  • β€’ Latency-sensitive conversational interfaces β€” TTFT is low relative to the reasoning variant; suitable for chat applications where waiting several seconds for reasoning is unacceptable
  • β€’ Long-context document processing β€” 2M-token window fits entire codebases, lengthy contract sets, or extended transcripts; pairs well with structured output for extraction tasks
  • β€’ Batch classification and extraction at scale β€” no reasoning overhead per call keeps cost predictable; suitable for categorisation, entity extraction, and templated transformations run at high volume
  • β€’ Tool-calling in multi-step workflows where another model handles reasoning β€” the non-reasoning variant is a reliable executor that handles tool dispatch and result formatting; cheaper than keeping a reasoning model active for every step
Reach for something else if
  • β€’ Complex multi-step reasoning tasks β€” without chain-of-thought, performance on hard math, scientific analysis, and Olympiad-level problems drops sharply; use xai/grok-4.20-reasoning or xai/grok-4.3 instead
  • β€’ Tasks where the same-price reasoning variant exists β€” at identical xAI list pricing ($2/$6 per 1M), most tasks benefit from enabling reasoning; the non-reasoning variant is only preferable when latency or per-step cost in a larger pipeline is the binding constraint
  • β€’ Cases requiring the latest xAI generation β€” Grok 4.3 (April 2026, 1M context, always-on reasoning, native video, $1.25/$2.50 per 1M) is the current xAI flagship and is now recommended for all new chat and coding integrations
  • β€’ Requests requiring video or audio modality β€” Grok 4.20 Non-Reasoning accepts text and image only; native video input was introduced in Grok 4.3; audio requires xAI's separate Voice API

How TheRouter serves this differently from the vendor

As the vendor operates it

xAI exposes grok-4.20-0309-non-reasoning as a native xAI API model with xAI pricing tiers, xAI rate limits, and the documented caveat that logprobs/top_logprobs are silently ignored for Grok 4.20 and newer.

On TheRouter

TheRouter serves the pinned xai/grok-4.20-non-reasoning route through its OpenAI-compatible chat endpoint with TheRouter pricing and routing metadata; unsupported media-generation, audio, and embeddings endpoints are not exposed for this model.

docs.x.ai β†—
Context Length
2M
Max Output
66K
Input Priceper 1M tokens
$1.35/ 1M tokens
Output Priceper 1M tokens
$2.70/ 1M tokens

Modalities

textimage→text

Capabilities

Vision

Pricing Breakdown

TypeRate
Prompts under 200K tokens
Input$1.35 / 1M tokens
Output$2.70 / 1M tokens
Cached input$1.35 / 1M tokens
Prompts of 200K tokens or more
Input$2.70 / 1M tokens
Output$5.40 / 1M tokens
Cached input$0.432 / 1M tokens

Tiered pricing: when a request's prompt (input, cached tokens included) is 200K tokens or more, every token in that request is billed at the long-context rates. This model publishes no separate cached-input price for the base group: cached tokens there are billed at the base input rate (that group's "Cached input" row shows that rate).

Supported Parameters

temperaturemax_tokenstop_ptoolstool_choiceresponse_formatstop

Specifications

Release date (public API)2026-03-09 / 2026-03-10unknown
API model slug (pinned)grok-4.20-0309-non-reasoningdocs.x.ai β†—verified
Reasoning modeOff (no chain-of-thought; for CoT use xai/grok-4.20-reasoning)verified
Context window2,000,000 tokensdocs.x.ai β†—verified
Max completion tokens65,536 tokensverified
Input modalitiesText, Image (jpg/png, up to 20 MiB per image)docs.x.ai β†—verified
Output modalitiesTextverified
xAI API pricing (current)$1.25 input / $2.50 output per 1M tokens (as of June 2026; TheRouter may apply a margin)docs.x.ai β†—verified
Function calling / toolsSupportedverified
Structured outputs (response_format)Supported (JSON mode)verified
logprobs / top_logprobsNot supported β€” silently ignored for grok-4.20 and newer (API accepts without error but returns no token-level log probabilities)docs.x.ai β†—verified
Knowledge cutoffNot publicly disclosed (listed as 'March 2026' in some aggregator metadata)unknown
LicenseProprietaryverified
Rate limits (xAI API)1,800 requests/min Β· 10,000,000 tokens/mindocs.x.ai β†—verified

Benchmarks

BenchmarkDistributionScoreSource
AA Intelligence Index
xAI's public model documentation does not disclose this benchmark score for grok-4.20-0309-non-reasoning. Keep this row unknown until xAI publishes an official value.
β€”Not publicly disclosedβ€”
AA Coding Index
β€”Not publicly disclosedβ€”
AA Agentic Index
β€”Not publicly disclosedβ€”
GPQA Diamond
xAI's public model documentation does not disclose this benchmark score for grok-4.20-0309-non-reasoning. Keep this row unknown until xAI publishes an official value.
β€”Not publicly disclosedβ€”
HLE (Humanity's Last Exam)
β€”Not publicly disclosedβ€”
SciCode
xAI's public model documentation does not disclose this benchmark score for grok-4.20-0309-non-reasoning. Keep this row unknown until xAI publishes an official value.
β€”Not publicly disclosedβ€”
τ²-Bench (Telecom)
xAI's public model documentation does not disclose this benchmark score for grok-4.20-0309-non-reasoning. Keep this row unknown until xAI publishes an official value.
β€”Not publicly disclosedβ€”
IFBench (Instruction Following)
β€”Not publicly disclosedβ€”
AA-LCR (Long-Context Reasoning)
xAI's public model documentation does not disclose this benchmark score for grok-4.20-0309-non-reasoning. Keep this row unknown until xAI publishes an official value.
β€”Not publicly disclosedβ€”
TerminalBench Hard
xAI's public model documentation does not disclose this benchmark score for grok-4.20-0309-non-reasoning. Keep this row unknown until xAI publishes an official value.
β€”Not publicly disclosedβ€”

API Usage Examples

Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.

cURL
curl https://api.therouter.ai/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $THE_ROUTER_API_KEY"   -d '{
    "model": "xai/grok-4.20-non-reasoning",
    "messages": [
      {"role": "user", "content": "Summarize the key points from this input."}
    ]
  }'

API guide

Full API reference β†’

Chat completion

cURL
curl https://api.therouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "xai/grok-4.20-non-reasoning",
    "messages": [{"role": "user", "content": "Summarise the key points in the following document:"}]
  }'

More from xai

xai/grok-4.20-reasoningxai/grok-4.20-multi-agentxai/grok-4.3xai/grok-4xai/grok-4-fast-reasoningxai/grok-4-fast-non-reasoningxai/grok-3xai/grok-3-mini

Similar models

Cross-provider sibling models
anthropic/claude-sonnet-4.6
google/gemini-2.5-flash
openai/gpt-4.1
deepseek/deepseek-v4

News & changes

2026-03-10

Grok 4.20 Beta launches with four API variants including non-reasoning mode

xAI released the Grok 4.20 Beta generation to the public API on March 9–10, 2026, introducing four concurrent variants: the default grok-4.20-beta, a non-reasoning speed variant (grok-4.20-0309-non-reasoning), a deep-reasoning variant (grok-4.20-0309-reasoning), and a multi-agent collaboration variant (grok-4.20-multi-agent-0309). All four share a 2M-token context window β€” the largest among Western frontier models at the time β€” and identical list pricing of $2/$6 per 1M tokens. The non-reasoning mode targets automated pipelines and tool-calling workflows where latency matters more than reasoning depth.

re-authored by TheRoutergrokipedia.com β†—

Frequently asked

What is the difference between grok-4.20-non-reasoning and grok-4.20-reasoning?

Both variants run the same underlying Grok 4.20 model weights. The non-reasoning variant (grok-4.20-0309-non-reasoning) skips the chain-of-thought step entirely, delivering lower latency and approximately 200 tokens/sec throughput. The reasoning variant (grok-4.20-0309-reasoning) activates extended chain-of-thought before generating the final answer, which significantly improves performance on complex math, scientific analysis, and multi-step logic at the cost of higher time-to-first-token. The benchmark gap is large: the non-reasoning variant scores 29.7 on the AA Intelligence Index versus a higher score for the reasoning variant. At the same xAI list price ($2/$6 per 1M tokens at launch), the non-reasoning variant is best suited for workloads where latency is the binding constraint β€” automated pipelines, tool dispatch, high-volume extraction β€” rather than maximum quality per response.

re-authored by TheRouter
Is the non-reasoning mode cheaper than the reasoning mode?

At xAI list pricing, both variants cost the same per input and output token ($2/$6 per 1M tokens at launch; xAI updated to $1.25/$2.50 for both in June 2026). The non-reasoning mode is effectively cheaper per completed task in latency-sensitive pipelines because it produces fewer total tokens (no reasoning tokens are billed) and completes faster, reducing wall-clock time. In batch workloads, both modes support 50% batch discounts on the xAI API. Check TheRouter's pricing page for current margins applied.

re-authored by TheRouter
When should I use Grok 4.20 Non-Reasoning vs Grok 4.3?

Grok 4.3 (April 2026) is xAI's current recommended model for all new integrations. It features always-on reasoning, 1M token context, native video input, and a lower price ($1.25/$2.50 per 1M at xAI list pricing). For most new workloads, xai/grok-4.3 is the better choice. The Grok 4.20 Non-Reasoning variant may still be preferable if you have an existing integration relying on the 2M-token context window (4.3 has 1M), or if your pipeline specifically benefits from the Grok 4.20's non-reasoning latency profile and you want a pinned stable model slug that will not be updated.

re-authored by TheRouter
Fact ledger β€” every claim on this page traces here
sourceURLretrieved
Release date (public API)β€”β€”unknown
API model slug (pinned)docs.x.ai β†—2026-06-09verified
Reasoning modeβ€”β€”verified
Context windowdocs.x.ai β†—2026-06-09verified
Max completion tokensβ€”β€”verified
Input modalitiesdocs.x.ai β†—2026-06-09verified
Output modalitiesβ€”β€”verified
xAI API pricing (current)docs.x.ai β†—2026-06-09verified
Function calling / toolsβ€”β€”verified
Structured outputs (response_format)β€”β€”verified
logprobs / top_logprobsdocs.x.ai β†—2026-06-09verified
Knowledge cutoffβ€”β€”unknown
Licenseβ€”β€”verified
Rate limits (xAI API)docs.x.ai β†—2026-06-09verified
AA Intelligence Indexdocs.x.ai β†—2026-08-01unknown
AA Coding Indexdocs.x.ai β†—2026-08-01unknown
AA Agentic Indexdocs.x.ai β†—2026-08-01unknown
GPQA Diamonddocs.x.ai β†—2026-08-01unknown
HLE (Humanity's Last Exam)docs.x.ai β†—2026-08-01unknown
SciCodedocs.x.ai β†—2026-08-01unknown
τ²-Bench (Telecom)docs.x.ai β†—2026-08-01unknown
IFBench (Instruction Following)docs.x.ai β†—2026-08-01unknown
AA-LCR (Long-Context Reasoning)docs.x.ai β†—2026-08-01unknown
TerminalBench Harddocs.x.ai β†—2026-08-01unknown
Grok 4.20 Beta launches with four API variants including non-reasoning modegrokipedia.com β†—2026-06-09verified
Help & contact