curl --request POST "https://gptproto.com/v1/chat/completions" \
--header "Authorization: Bearer $GPTPROTO_API_KEY" \
--header "Content-Type: application/json" \
--data '{
"model": "grok-4.7",
"messages": [
{
"role": "user",
"content": "Hello"
}
]
}'Chat, Coding-Agenten und Dokumentenarbeit. Preise pro 1 Mio. Tokens – Eingabe, gecachte Eingabe und Ausgabe werden separat abgerechnet. GPTProto liegt 40% unter den offiziellen Preisen.
| Szenario | Grok Liste | OpenRouter | GPTProto | Du sparst / Monat |
|---|---|---|---|---|
| Persönlich10 Mio. Tokens / Monat (4.8 Mio. im Cache) | $20.80 | $21.94 | $12.48 | −$8.32≈ $99.84 / Jahr |
| Team100 Mio. Tokens / Monat (48 Mio. im Cache) | $208.00 | $219.44 | $124.80 | −$83.20≈ $998.40 / Jahr |
| Business500 Mio. Tokens / Monat (240 Mio. im Cache) | $1040.00 | $1097.20 | $624.00 | −$416.00≈ $4992.00 / Jahr |
Grok · 500K context · $1.2/M input · $3.6/M output
Bytedance · 262K context · $0.0971/M input · $0.9714/M output
Bytedance · 262K context · $0.0971/M input · $0.9714/M output
Die Grok 4.7 API bietet Entwicklern Zugriff auf das Modell von SpaceXAI für Softwareentwicklung, agentische Aufgaben und Wissensarbeit. Es verarbeitet Text- und Bildeingaben, gibt Text aus, unterstützt Function Calling und strukturierte Ausgaben und bietet den Reasoning-Aufwand low, medium, high und xhigh. Mit GPTProto können Teams das Modell zusammen mit anderen Modellen über ein einziges Konto und Guthaben testen.
500K multimodaler Kontext
Verarbeiten Sie umfangreiche Repositories, technische Dokumente, UI-Screenshots und Diagramme in einer Anfrage. Das Kontextfenster umfasst sowohl Prompt- als auch generierte Tokens. Planen Sie daher ausreichend Kapazität für die erwartete Antwort ein.
Reasoning von Low bis xHigh
Wählen Sie den Reasoning-Aufwand low, medium, high oder xhigh. Beginnen Sie bei Routineaufgaben unter xhigh und erhöhen Sie den Aufwand nur dann, wenn eine gründlichere Überprüfung die Aufgabenerledigung verbessert.
Lang laufende Coding-Aufgaben
Nutzen Sie Grok 4.7 für Änderungen an mehreren Dateien, Debugging, Repository-Analysen, Testplanung und wiederholte Plan-Ausführen-Prüfen-Schleifen, bei denen das Modell über viele Schritte hinweg den Zustand beibehalten muss.
Tool-Aufrufe und strukturierte Ausgabe
Verbinden Sie das Modell mit Anwendungsfunktionen und fordern Sie JSON an, das einem Schema entspricht. Diese Funktionen helfen Agenten, validierte Daten zwischen Tools weiterzugeben, statt sich auf unstrukturierte Prosa zu verlassen.
Grok 4.7 is a text-output model released by SpaceXAI on September 21, 2026. It uses a larger base model than Grok 4.6 and received a longer reinforcement-learning run weighted toward difficult tasks that may take hours to complete.
The model accepts text and images within a 500,000-token context window. SpaceXAI does not publish a separate model-level text output limit, but input plus generated output must fit within the context window. Gateway timeouts and account limits still apply.
Developers can use function calling, structured outputs, and configurable reasoning. Web or X search must be explicitly enabled for current information; the published knowledge cutoff is May 2026.
Through GPTProto, the same account and balance can be used for Grok 4.7 and 200+ other models, simplifying comparisons and fallbacks across providers.
| Specification | Grok 4.7 |
|---|---|
| Provider | SpaceXAI (xAI) |
| Release date | September 21, 2026 |
| GPTProto model string | grok-4.7 |
| Input modalities | Text and image |
| Output modality | Text |
| Context window | 500,000 tokens, including input and generated output |
| Published text output limit | No separate model-level limit listed by SpaceXAI |
| Reasoning effort | low, medium, high, xhigh; default high in the official API |
| Agent controls | Function calling and structured outputs |
| Batch API | Not supported in the official SpaceXAI API at launch |
| Knowledge cutoff | May 2026 |
Repository-scale coding: Let an agent inspect interfaces, tests, configuration, and dependencies before proposing a change. Request a plan, patch summary, test evidence, and unresolved risks as separate fields.
Long-horizon agentic work: Use function calls for search, execution, validation, and retries. Keep permissions and completion criteria in the surrounding agent harness.
Visual technical analysis: Supply UI screenshots, charts, or diagrams with text instructions. Use structured output when downstream code needs coordinates, labels, issue categories, or confidence fields.
Professional knowledge work: Draft reports or presentations from large source sets. For high-stakes use, keep citations, retrieval, and human approval outside the model response.
Grok 4.7 keeps Grok 4.6's 500K context, modalities, official token rates, and reasoning levels. The change is behavioral: a larger base model, longer reinforcement learning, stronger self-checking, and better management of extended work.
| Decision factor | Grok 4.7 | Grok 4.6 |
|---|---|---|
| Context window | 500K tokens | 500K tokens |
| Input / output | Text and image / text | Text and image / text |
| Reasoning levels | Low, medium, high, xhigh | Low, medium, high, xhigh |
| Official price below 200K prompt tokens | $2 input / $0.50 cached input / $6 output per 1M tokens | $2 input / $0.50 cached input / $6 output per 1M tokens |
| CursorBench 4.0, provider-reported | 46.3% at xhigh | 40.4% at high |
| DeepSWE v1.1, provider-reported | 71.0% at high | 65.2% at high |
| Terminal-Bench 4.0, provider-reported | 38.0% at xhigh | 20.3% at high |
| Best fit | New coding and knowledge-work deployments; harder multi-step tasks | Existing integrations that still need regression testing before migration |
Benchmark results above are reported by SpaceXAI and were not independently reproduced by GPTProto. Reasoning settings differ between some rows, so treat them as launch evidence rather than a universal ranking.
No single benchmark answers every comparison in this group. SpaceXAI reports 46.3% for Grok 4.7 on CursorBench 4.0, versus 41.7% for GPT-5.6 Sol and 51.8% for Claude Fable 5.1. On DeepSWE v1.1, it reports 71.0%, 72.7%, and 70.0%, respectively. The winner changes with the workload.
SpaceXAI also includes GPT-6 Astra in knowledge-work charts, but those results should not be merged with the coding table above. For Grok 4.7 vs GPT-6 Astra, Claude Fable 5.1 vs Grok 4.7, Qwen3.8-Max-0902 vs Grok 4.7, or Grok 4.7 vs Gemini 4, use identical prompts and tools, then record completion, retries, tokens, latency, and cost.
Higher effort can improve difficult tasks but may increase latency and reasoning-token use. Choose the lowest level that reliably completes the job.
| Workload | Suggested starting effort | Evaluation focus |
|---|---|---|
| Classification, extraction, or simple code edits | low |
Schema validity, latency, cost |
| Multi-file feature work or document synthesis | medium |
Cross-file consistency, missed requirements |
| Difficult debugging or multi-tool agents | high |
Task completion, tool errors, recovery behavior |
| Hard architecture, deep investigation, or failed high-effort runs | xhigh |
Added success rate versus extra tokens and time |
For long tasks, checkpoint state, set tool timeouts, cap retries, validate structured outputs, and log tokens by step. This is more informative than comparing list prices alone.
Start a Grok 4.6 migration with a model-string change, not a full prompt rewrite. Keep the old version available and compare both on the same task set.
Confirm the GPTProto model string shown in the page's Quick Start panel before changing application code.
Re-run regression tasks at the same reasoning effort. Do not compare Grok 4.6 high against Grok 4.7 xhigh unless the extra token use is part of the test.
Check tool-call arguments, structured JSON, stop behavior, output length, latency, and total cost—not just answer quality.
Review prompts that assume a fixed output cap. SpaceXAI lists no separate text output limit, but the full request still has to fit within 500K tokens.
Keep a fallback until Grok 4.7 passes the acceptance criteria for your real repositories, tools, and documents.
When switching from direct xAI access, use the endpoint and authentication pattern in GPTProto's fixed Quick Start section. Do not copy the official xAI base URL into a GPTProto request.
Anleitungen, Vergleiche und Updates zu diesem Modell.
Alle Artikel
GLM-5.3-FlashX verspricht bis zu 200 Tokens/s – kostet aber 2,5-mal so viel wie Flash. Erfahre mehr über Preise, Eignung fürs Programmieren, Geschwindigkeitsgrenzen und wann sich das Upgrade lohnt.
Vergleiche die sechs besten LLM-API-Anbieter 2026 hinsichtlich Modellzugriff, Preisstruktur, Geschwindigkeit, Kompatibilität, Governance und Developer Experience.
Was ist Jev? Erfahre, wie das Entscheidungsmodell von TypeSafe AI typisierte Wahrscheinlichkeiten zurückgibt, worin es sich von LLMs unterscheidet und wie es sich für Routing- und Agenten-Workflows eignet.
Vergleiche GPTProto, Supabase, Stripe, Clerk und Resend, um ein schlankes, API-first AI-SaaS-MVP mit geringeren anfänglichen Infrastrukturkosten zu entwickeln.
