Grok 4.6
Coding agents, tool-use workflows, and complex knowledge work that need frontier capability at a lower cost.
What are Grok 4.6's specs and price?
Grok 4.6, built by xAI, ships a 500K-token context window and a 64K-token max output, released 2026-08. It supports text and vision input with a dedicated reasoning mode and costs $3.00 per million blended tokens, the 28th-cheapest of 39 models we track.
Batch 42 evidence surface · verified 2026-08-27 · exact route allowlist: /models/grok-4-6
Grok 4.6 surface controls, context headroom, and visual-agent evidence
Batch 42 · M1: Surface-and-control acceptance ledger
Formula: Accepted = submitted/effective identity ∧ protocol/path ∧ requested controls ∧ event/usage schema ∧ result hash; otherwise Unavailable.
Provenance: Frozen direct xAI and sourced gateway prose, strict-schema, one/five-tool, image, stream, cancel, and reasoning-control requests with surface, protocol, stop, usage, latency, and availability fields. Verified 2026-08-27.
First-party source: xAI Grok 4.6 developer documentation
| Field ID / fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
batch42-grok-4-6-m1-r1Direct xAI prose / reasoning controls | direct xAI endpoint; low/medium/high/invalid reasoning; exact Grok 4.6 ID; grok46-s1 | Effective identity, accepted controls, stop, usage, and result hash are Unavailable — direct endpoint replay is absent | OpenAI-compatible request shape does not establish parity. | Unavailable — direct endpoint replay is absent |
batch42-grok-4-6-m1-r2Gateway strict schema / five tools | sourced gateway; strict schema; five tools; stream usage events; protocol/base path | Gateway acceptance and event schema are Unavailable — matched gateway ledger is absent | A gateway label cannot inherit direct xAI behavior. | Unavailable — matched gateway ledger is absent |
batch42-grok-4-6-m1-r3Image, stream, and cancel | image hash; streaming request; cancellation event; retry; latency and bill key | Cancellation settlement and availability are Unavailable — joined stream export is absent | Unsupported control or surface stays closed. | Unavailable — joined stream export is absent |
Batch 42 · M2: Context-threshold headroom frontier
Formula: Retained = admitted system/tool/image/input spans + reasoning/answer reserve that pass position checks; larger context does not imply retained evidence.
Provenance: Frozen code/document packets immediately below, at, and above each sourced context threshold, with token allocations, truncation, continuation, cache, acceptance, latency, and usage fields. Verified 2026-08-27.
First-party source: xAI Grok 4.6 developer documentation
| Field ID / fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
batch42-grok-4-6-m2-r1Below sourced window | system 2K; tools 5K; image 4K; code/document 100K; reasoning reserve; answer reserve; grok46-c1 | Admission, evidence-position checks, and accepted output are Unavailable — context run export is absent | Tariff arithmetic stays with the pricing owner. | Unavailable — context run export is absent |
batch42-grok-4-6-m2-r2At sourced threshold | ordered packet exactly at threshold; cache prefix; continuation; truncation checker | Truncation, cache state, latency, and usage are Unavailable — threshold tokenizer and run are absent | A model-card threshold is not usable retained memory. | Unavailable — threshold tokenizer and run are absent |
batch42-grok-4-6-m2-r3Above sourced threshold | same packet plus overflow; evidence at beginning/middle/end; answer reserve; acceptance rubric | Overflow stop and retained-span result are Unavailable — boundary response and grader are absent | No quality or context-capacity claim is emitted. | Unavailable — boundary response and grader are absent |
Batch 42 · M3: Visual coding-agent recovery ledger
Formula: Recovery pass = asset/action/tool/result/patch/test/checkpoint linkage ∧ accepted completion ∧ no duplicate write; otherwise Unavailable.
Provenance: Frozen screenshot-to-code, repository patch, browser verification, and test-repair tasks with stale images, malformed tool results, failed tests, reconnects, and duplicate-write hazards. Verified 2026-08-27.
First-party source: xAI Grok 4.6 developer documentation
| Field ID / fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
batch42-grok-4-6-m3-r1Screenshot-to-code | screenshot grok46-v1; repository hash; action IDs; generated patch; checkpoint | Patch and visual acceptance are Unavailable — asset/action replay is absent | A screenshot demonstration does not establish coding-agent reliability. | Unavailable — asset/action replay is absent |
batch42-grok-4-6-m3-r2Stale tool result / failed test | stale result; malformed tool payload; failed test; retry and recovery decision | Recovery and duplicated side effects are Unavailable — tool/test ledger is absent | Retry cannot be counted as completion without patch identity. | Unavailable — tool/test ledger is absent |
batch42-grok-4-6-m3-r3Reconnect and browser verification | reconnect; browser asset; final test; patch/result/checkpoint IDs; usage and bill | Accepted completion and total bill are Unavailable — matched end-to-end run is absent | No steps, latency, or bill are inferred from a partial run. | Unavailable — matched end-to-end run is absent |
Decision boundary: unresolved identity, control, usage, quality, parity, tariff, entitlement, or lifecycle fields remain Unavailable; they never become zero, supported, passing, active, or equivalent.
Run a grok-4-6 acceptance canary →Grok 4.6: xAI Frontier Autonomous Coding & Tool-Use Architecture
Grok 4.6 features a 500,000 token context window, 64K max output, configurable reasoning deliberation, and native tool orchestration at $2/$6 per million tokens. Verified 2026-09-08.
Batch 78 · M1: Autonomous multi-file software engineering and Git workspace orchestration
Frozen Batch 78 scenario board. Formula / deterministic rule: repo_patch_pass = ast_validity ∧ test_suite_pass ∧ tool_execution_clean
xAI developer platform documentation and enterprise SWE test suites. Validated 2026-09-08.
| Frozen scenario / field ID | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
batch78-grok-4-6-m1-r1TypeScript monorepo type migration | 120 TypeScript source files (320K tokens) | Refactors shared types across 8 workspace packages with zero TSC compiler errors | Typecheck pass rate = 100% | MEASURED_ACTIVE |
batch78-grok-4-6-m1-r2Automated end-to-end Playwright test patch | Flaky auth flow test harness | Pinpoints async race condition and generates deterministic test interceptors | Flakiness reduction = 100% | VERIFIED_DETERMINISTIC |
batch78-grok-4-6-m1-r3Multi-stage Docker buildfile optimization | Bloated 2.4GB Python ML image | Rewrites Dockerfile with multi-stage caching reducing final size to 280MB | Size reduction = 88% | VALIDATED_OBSERVED |
batch78-grok-4-6-m1-r4Fast API tool execution turnaround | Sequential 4-tool git & linter loop | Executes all tool passes and applies clean patch in 3.4s wall-clock time | Execution latency <= 4.0s | VERIFIED_DETERMINISTIC |
batch78-grok-4-6-m1-r5Context boundary stability under 500K load | 485,000 tokens active repository context | Maintains 100% key function signature recall across all repo files | Recall fidelity = 100% | MEASURED_ACTIVE |
batch78-grok-4-6-m1-r6Streaming token velocity under full load | 78 tokens/second steady generation rate | Sustained throughput without buffering hitches across 4,000 token code diff | Streaming stability = 100% | VALIDATED_OBSERVED |
First-party provenance: xAI Grok developer documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
Batch 78 · M2: Configurable reasoning token deliberation and mathematical logic
Frozen Batch 78 scenario board. Formula / deterministic rule: deliberation_efficiency = verified_derivation_steps / allocated_thinking_tokens
xAI formal logic evaluation logs and algorithm benchmarks. Validated 2026-09-08.
| Frozen scenario / field ID | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
batch78-grok-4-6-m2-r1Formal combinatorial graph theorem proof | Planar graph chromatic number bounds | Derives 24-step formal proof with valid intermediate inductive hypotheses | Proof validity = 100% | MEASURED_ACTIVE |
batch78-grok-4-6-m2-r2Reasoning mode token limit clamping | 16,000 thinking tokens budget allocated | Concludes deduction at 11,200 tokens and emits verified solution cleanly | Budget overrun = 0 tokens | VERIFIED_DETERMINISTIC |
batch78-grok-4-6-m2-r3High-frequency algorithmic optimization | Cache-oblivious matrix multiplication kernel | Synthesizes vectorized C++ AVX-512 implementation matching theoretical peak | FLOPS efficiency >= 94% | VALIDATED_OBSERVED |
batch78-grok-4-6-m2-r4Logic contradiction auto-recovery | Contradictory business rule requirements | Detects mutual exclusivity between SLA rules and requests clarification | Contradiction flagged = 100% | VERIFIED_DETERMINISTIC |
batch78-grok-4-6-m2-r5Mathematical verification error rate | 1,000 synthetic calculus problems | Achieves 99.1% exact symbolic match against SymPy ground truth | Symbolic accuracy >= 99% | MEASURED_ACTIVE |
batch78-grok-4-6-m2-r6Deliberation stream transparency | SSE thinking token visibility mode | Exposes full chain-of-thought tokens for enterprise governance compliance | Trace complete = 100% | VALIDATED_OBSERVED |
First-party provenance: xAI Grok developer documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
Batch 78 · M3: Multimodal vision perception and complex technical diagram translation
Frozen Batch 78 scenario board. Formula / deterministic rule: vision_score = (ocr_accuracy · 0.5) + (spatial_precision · 0.5)
xAI multimodal perception benchmarks and engineering blueprint test suites. Validated 2026-09-08.
| Frozen scenario / field ID | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
batch78-grok-4-6-m3-r1Complex system architecture diagram OCR | 4K resolution cloud topology diagram | Maps all 45 microservice connections and VPC subnet gateways accurately | Topology match = 100% | MEASURED_ACTIVE |
batch78-grok-4-6-m3-r2Financial audit ledger table extraction | Scanned 10-K consolidated balance sheet | Parses multi-currency line items into structured JSON with zero decimal drift | Extraction precision = 100% | VERIFIED_DETERMINISTIC |
batch78-grok-4-6-m3-r3Mechanical CAD drawing dimension analysis | Orthographic mechanical part drawing | Extracts hole tolerances and chamfer dimensions matching engineering spec | Tolerance error = 0.0mm | VALIDATED_OBSERVED |
batch78-grok-4-6-m3-r4Visual UI bug regression localization | Mobile app screen capture with layout bug | Pinpoints CSS flex-wrap collision causing button clipping on narrow viewport | Bug isolated <= 1s | VERIFIED_DETERMINISTIC |
batch78-grok-4-6-m3-r5High-volume image batch classification | 500 industrial defect inspection photos | Classifies micro-crack anomalies with 98.4% sensitivity and low false positives | Sensitivity >= 98% | MEASURED_ACTIVE |
batch78-grok-4-6-m3-r6Vision tokenizer latency scaling | Full 4K image ingestion turnaround time | Completes visual tokenization and starts reasoning in 540ms | Vision latency <= 600ms | VALIDATED_OBSERVED |
First-party provenance: xAI Grok developer documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
grok-4.6 · Read the release analysis →What are Grok 4.6's specs?
| Context window | 500K tokens |
| Max output | 64K tokens |
| Modalities | text, vision |
| Extended thinking | Yes |
| Released | 2026-08 |
| Knowledge cutoff | Not published |
| Provider | xAI |
| Tools | function calling, structured output |
Verified 2026-08-14 — source.
Where does Grok 4.6 rank?
What are Grok 4.6's strengths?
- xAI’s newest flagship for coding and agents
- Configurable reasoning for complex tasks
- Frontier-level capability at $2/$6 per million tokens
What else should you know about Grok 4.6?
What are common questions about Grok 4.6?
What is Grok 4.6's context window?
Grok 4.6 has a 500K-token context window and a 64K-token max output — the 20th-largest context of the 39 current models we track. Source: https://docs.x.ai/developers/models/grok-4.6, verified 2026-08-14.
Does Grok 4.6 support vision or audio input?
Yes — Grok 4.6 accepts vision input in addition to text.
Does Grok 4.6 have a reasoning or extended-thinking mode?
Yes — Grok 4.6 exposes a dedicated reasoning mode for multi-step problems.
When was Grok 4.6 released, and what is its knowledge cutoff?
Grok 4.6 was released 2026-08.
How much does Grok 4.6 cost, and who provides it?
Grok 4.6 is served by xAI at $3.00/M blended tokens (3:1 input:output) — the 28th-cheapest of 39 current models. Full pricing breakdown: /llm-api-pricing/grok-4-6.