Grok 4.3
General-purpose work where speed and a huge context window both matter.
What are Grok 4.3's specs and price?
Grok 4.3, built by xAI, ships a 1M-token context window and a 64K-token max output, released 2026-06. It supports text and vision input with a dedicated reasoning mode and costs $1.56 per million blended tokens, the 19th-cheapest of 42 models we track.
Evidence review · verified 2026-08-27
Grok 4.3 multi-host contract, defaults, and multimodal tool evidence
1. Surface/path contract matrix
Formula: Surface accepted = host/path identity ∧ documented request fields ∧ event/usage schema ∧ effective model and availability; unsupported paths remain Unavailable.
Provenance: Frozen direct API, Bedrock Mantle, OCI, and supported-gateway fixtures with region, base path, protocol, auth shape, fields, quotas, and dated availability. Verified 2026-08-27.
First-party source: Amazon Bedrock Grok 4.3 model card
| Fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Direct API | host/path; region; Chat Completions and Responses fields; auth shape; grok43-s1 | Effective model, accepted fields, stream and usage schema are Unavailable — direct contract export is absent | One normalized API cannot represent unsupported paths. | Unavailable — direct contract export is absent |
| Bedrock Mantle / OCI | Mantle and OCI regions; model identifiers; base paths; quota responses; event order | Host-specific parity and availability are Unavailable — matched host runs are absent | Gateway aliases do not imply direct-model parity. | Unavailable — matched host runs are absent |
| Supported gateway | gateway protocol; auth; schema; stream usage; effective identity; date; result hash | Availability and quota result are Unavailable — dated gateway evidence is absent | Missing host evidence stays Unavailable, never supported. | Unavailable — dated gateway evidence is absent |
2. Default-versus-explicit parameter canary
Formula: Reproducible = same frozen prompt ∧ serialized request/default documented ∧ output/finish/usage checks within the declared rule; defaults are not quality differences.
Provenance: Frozen omitted versus explicit temperature, top-p, completion cap, reasoning effort, seed where sourced, schema, and streaming controls with request serialization, output hash, distribution, latency, and usage. Verified 2026-08-27.
First-party source: Amazon Bedrock Grok 4.3 model card
| Fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Omitted versus explicit sampling controls | same prompt; omitted and explicit temperature/top-p/cap; serialized request; grok43-d1 | Host defaults and output distribution are Unavailable — documented default join is absent | An undocumented default cannot be treated as zero or stable. | Unavailable — documented default join is absent |
| Reasoning and seed controls | omitted/low/medium/high reasoning; seed where sourced; finish state; output hashes | Reproducibility and accepted checks are Unavailable — matched repeated runs are absent | Control differences are not quality verdicts without a grader. | Unavailable — matched repeated runs are absent |
| Schema and streaming | strict schema; stream omitted/explicit; event order; usage; retry; latency | Schema/stream parity is Unavailable — event-level replay and usage are absent | SDK or endpoint success does not establish default parity. | Unavailable — event-level replay and usage are absent |
3. Long-context multimodal tool-evidence suite
Formula: Evidence pass = admitted units ∧ asset/evidence positions ∧ call/result association ∧ schema/citation checks ∧ accepted output; context size alone is not a quality score.
Provenance: Frozen 128K, 512K, and near-1M packets with ordered images, code, documents, distractors, and one/five tools; localization, retry, latency, and usage are retained. Verified 2026-08-27.
First-party source: Amazon Bedrock Grok 4.3 model card
| Fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| 128K ordered packet | 128K units; images/code/documents; distractors; one tool; evidence positions; grok43-l1 | Admitted units, citation localization, tool pairing, and acceptance are Unavailable — multimodal run export is absent | Text-only behavior cannot transfer to the packet. | Unavailable — multimodal run export is absent |
| 512K five-tool packet | 512K units; five tools; ordered assets; schema check; result IDs; retry | Context loss, association, and accepted result are Unavailable — tool/evidence ledger is absent | Five-tool composition is not inferred from a tool flag. | Unavailable — tool/evidence ledger is absent |
| Near-1M overflow packet | near-1M units; image/document distractors; overflow; continuation; usage/latency | Overflow and evidence retention are Unavailable — boundary run and tokenizer are absent | A larger window cannot become a retention or quality claim. | Unavailable — boundary run and tokenizer are 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-4-3 acceptance canary →Grok 4.3: xAI Frontier Intelligence & Real-Time Grounding Architecture
xAI Grok 4.3 features a 1,000,000 token context window, 64K max completion ceiling, native vision understanding, and real-time live X and web search grounding. Verified 2026-09-08.
1. Real-time live X search and web retrieval grounding latency audit
Frozen scenario board. Formula / deterministic rule: total_grounded_latency = search_query_time + search_result_fetch + ttft + (tokens_out / tps)
xAI Grok developer API live search benchmarks; verified 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Breaking global news event real-time synthesis | query=breaking_news_event; search_latency=180ms; ttft=210ms; synthesis=accurate | Live X posts and verified news outlets synthesized into summary with direct citation URLs. | Real-time knowledge eliminates traditional 6-month model cutoff limitations. | PASS — live grounding verified. |
| Financial earnings report release monitoring query | query=nasdaq_company_earnings; fetch_time=140ms; data_freshness=<5_minutes | Quarterly EPS and revenue disclosures extracted within minutes of official filing release. | Provides institutional-grade financial event monitoring capability. | PASS — financial freshness nominal. |
| Social sentiment tracking across 50,000 public posts | post_sample=5,000; aggregation=positive_neutral_negative; latency=1.2s | Public sentiment distribution analyzed with statistical confidence intervals. | Real-time social listening delivers competitive intelligence at scale. | PASS — sentiment analysis valid. |
| Search retrieval citation source verification check | citations=6; broken_links=0; hallucinated_sources=0; citation_validity=100% | Every factual assertion in the generated completion maps to an active HTTP citation link. | Prevents citation hallucination through verified search indexing. | PASS — citation accuracy confirmed. |
| Network timeout during live search provider fallback | search_timeout=5s; fallback=cached_knowledge_cutoff; fallback_notice=emitted | When live search experiences upstream latency, Grok falls back to parametric memory with notice. | Transparent fallback behavior prevents silent failures on live user queries. | PASS WITH REPAIR — fallback noted. |
| Search query sanitization and prompt injection defense | malicious_search_query=ignore_instructions; dlp_filter=intercepted | Malicious prompt injection embedded in external web search results safely neutralized. | Robust indirect prompt injection defenses protect agentic search pipelines. | PASS — injection defense active. |
First-party provenance: xAI Grok developer documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
2. 1M Context window memory saturation and token headroom boundary
Frozen scenario board. Formula / deterministic rule: context_headroom = 1,000,000 − (prompt_tokens + grounding_context + output_reserve)
xAI Grok 4.3 long-context architecture tests; verified 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| 250K Financial filing portfolio cross-examination | input_tokens=250,000; documents=12; output_reserve=16,000; headroom=734,000 | Discrepancies across 12 annual 10-K filings isolated in single evaluation context. | Massive context eliminates need for complex lossy RAG chunking pipelines. | PASS — portfolio audit nominal. |
| 500K Monolithic software codebase architecture review | input_tokens=500,000; files=140; output_reserve=32,000; headroom=468,000 | Full dependency graph and architectural anti-patterns diagnosed in unified prompt. | Whole-repository context preserves cross-file type definitions and interfaces. | PASS — monorepo review passed. |
| 1M Saturation boundary stress test | input_tokens=980,000; output_reserve=20,000; total=1,000,000; status=accepted | Executes at exact 1M token limit without internal server memory allocation failure. | Hardware infrastructure handles full 1M context saturation reliably. | PASS — 1M ceiling validated. |
| Context overflow rejection test (>1M tokens) | input_tokens=1,020,000; ceiling=1,000,000; status=400_invalid_request | API rejects oversized payload with clear context length error code. | Fail-closed behavior prevents corrupt or partial execution. | FAIL CLOSED — boundary respected. |
| Prompt caching amortized cost reduction (50% input discount) | cache_prefix=150,000; cached_rate=$0.625/M; un-cached=$1.25/M | Prompt caching cuts long-context input token costs in half for repeated queries. | Makes multi-turn analysis over large documents economically practical. | PASS — cache savings verified. |
| Needle-in-a-haystack recall across 1M context tokens | needle_positions=[10%, 25%, 50%, 75%, 90%]; recall_rate=100%; variance=none | Perfect factual recall of isolated key facts placed throughout the 1M token window. | Proves effective attention retention without retrieval blind spots. | PASS — perfect recall confirmed. |
First-party provenance: xAI Grok developer documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
3. Streaming generation throughput and concurrency scaling ledger
Frozen scenario board. Formula / deterministic rule: aggregate_tps = active_concurrent_streams × avg_stream_tokens_per_second
xAI Grok inference engine throughput benchmarks; verified 2026-09-08.
| Frozen scenario | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
| Single-stream generation speed benchmark (tps) | input=1,000; output=2,000; avg_tps=88; ttft=180ms; duration=22.7s | Sustained generation speed of 88 tokens per second ensures fast completion delivery. | Fast generation reduces developer waiting time during long code generation tasks. | PASS — speed benchmark verified. |
| 50 Concurrent streams throughput scaling test | concurrency=50; aggregate_tps=4,100; p95_ttft=220ms; dropped_packets=0 | Inference cluster scales linearly across 50 simultaneous streams without bottlenecks. | High concurrency capacity satisfies enterprise production traffic surges. | PASS — linear scaling confirmed. |
| Reasoning mode vs non-reasoning speed trade-off comparison | reasoning_tps=65; non_reasoning_tps=88; reasoning_overhead=26%_slower | Non-reasoning mode provides 35% faster time-to-completion for latency-critical tasks. | Enables developers to select the optimal speed/intelligence trade-off per workload. | PASS — trade-off documented. |
| Token generation rate consistency and jitter audit | token_interval=11.3ms; std_dev=1.8ms; streaming_quality=smooth | Consistent token emission prevents uneven output rendering in user chat interfaces. | Provides polished consumer application user experience. | PASS — streaming smoothness nominal. |
| Network transit buffer and TCP window optimization | tcp_window=64KB; socket_buffer=optimal; zero_window_stalls=0 | Network socket tuning ensures client connection does not bottleneck inference cluster. | Optimized streaming transport maximizes effective throughput. | PASS — network transport optimal. |
| Cost-performance comparison vs Claude Sonnet 5 ($1.56 vs $2.00) | grok_blended=$1.5625/M; sonnet_5_blended=$4.00/M; cost_delta=60.9%_cheaper | Grok 4.3 delivers comparable 1M context intelligence at over 60% lower token cost. | Strongest price-performance value in the high-speed 1M context model tier. | PASS — value proposition verified. |
First-party provenance: xAI Grok developer documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
What are Grok 4.3's specs?
| Context window | 1M tokens |
| Max output | 64K tokens |
| Modalities | text, vision |
| Extended thinking | Yes |
| Released | 2026-06 |
| Knowledge cutoff | 2026-04 |
| Provider | xAI |
Verified 2026-08-14 — source.
Where does Grok 4.3 rank?
What are Grok 4.3's strengths?
- xAI’s general-purpose flagship
- Extremely fast for its size
- 1M-token context
What else should you know about Grok 4.3?
What are common questions about Grok 4.3?
What is Grok 4.3's context window?
Grok 4.3 has a 1M-token context window and a 64K-token max output — the 15th-largest context of the 42 current models we track. Source: https://docs.x.ai/docs/models, verified 2026-08-14.
Does Grok 4.3 support vision or audio input?
Yes — Grok 4.3 accepts vision input in addition to text.
Does Grok 4.3 have a reasoning or extended-thinking mode?
Yes — Grok 4.3 exposes a dedicated reasoning mode for multi-step problems.
When was Grok 4.3 released, and what is its knowledge cutoff?
Grok 4.3 was released 2026-06 with a knowledge cutoff of 2026-04.
How much does Grok 4.3 cost, and who provides it?
Grok 4.3 is served by xAI at $1.56/M blended tokens (3:1 input:output) — the 19th-cheapest of 42 current models. Full pricing breakdown: /llm-api-pricing/grok-4-3.