Grok 4.5: Opus-Class Capability at Workhorse Pricing

SpaceXAI just opened Grok 4.5 to the public. On X, Elon Musk called it an Opus-class model that is faster, more token-efficient, and lower cost — with an internal read that puts it roughly next to Anthropic Opus 4.7 on capability while staying much cheaper to run. For builders, the interesting part is not the slogan. It is the combination of a 500K context window, strong coding and agent tooling, and list pricing at $2 / $6 per million tokens.

This post summarizes what shipped on July 8–9, 2026, what the official docs say, and how to try xai/grok-4.5 through an OpenAI-compatible gateway without rewriting your stack.

What Grok 4.5 is

Grok 4.5 is SpaceXAI's frontier text model aimed at coding, agentic loops, and knowledge work. Official docs list the API name as grok-4.5, with aliases grok-4.5-latest and grok-build-latest. It accepts text and image input, returns text, and supports function calling, structured outputs, and adjustable reasoning effort (low / medium / high, default high).

SpaceXAI positions the release as a workhorse for app building, research, writing, and routine office automation — not a niche research demo. Training ran in Memphis on large GPU clusters with new science, engineering, and math datasets. Coverage of the launch also notes supplemental training with Cursor-oriented developer data, which matches how the company is shipping the model as the default in Grok Build and inside Cursor on all plans.

Specs that matter for production

Property	Grok 4.5
API model ID	grok-4.5
Aliases	grok-4.5-latest, grok-build-latest
Context window	500,000 tokens
Modalities	Text + image → text
Reasoning	Low / medium / high (default high)
Tools	Function calling, web search, X search, code execution
Input price	$2.00 / 1M tokens
Cached input	$0.50 / 1M tokens
Output price	$6.00 / 1M tokens
Serving (vendor claim)	~80 tokens/sec
Regions (API console)	us-east-1, us-west-2 (EU console access expected later in July)

Two operational details from the docs are easy to miss. First, set a stable conversation routing key (prompt_cache_key on the Responses API, or the Chat Completions header SpaceXAI documents) so prompt-cache hits land on a warm server — otherwise you often pay full input price. Second, long agent loops benefit from context compaction; treat tool-heavy workloads as a first-class design problem, not an afterthought.

Why the price/performance pitch is loud

Musk's X posts framed Grok 4.5 as Opus-class but faster and cheaper. SpaceXAI's own materials lean on token efficiency: roughly 2× better than leading peers in their framing, with SWE Bench Pro numbers that show far fewer average output tokens than Opus 4.8 (max) on the same harness. Fewer output tokens usually means lower bill and lower latency per completed task — even before you compare sticker prices.

Sticker prices alone already look aggressive versus frontier peers. At $2 input / $6 output, Grok 4.5 undercuts typical Opus-tier list pricing by a wide margin. That only matters if quality holds on your evals. Vendor charts are mixed: Fable (max) still leads several coding benches in SpaceXAI's published comparison, while Grok 4.5 stays close on Terminal Bench 2.1 and posts strong legal-agent results (Harvey Legal Agent Benchmark #1 in SpaceXAI's reporting). Treat public benches as a starting filter, then run a small golden set of your own agent traces before you move production traffic.

Where it fits in a multi-model stack

Use Grok 4.5 when you want:

• Agentic coding and repair loops that burn many tool calls
• Long-context knowledge work inside a 500K window without jumping to a 1M-class bill
• A default model for Grok Build / Cursor-style workflows that still needs an API path
• Cost-sensitive production routes that previously defaulted to Opus-class models for "just in case" quality

Keep a stronger or differently specialized model on standby for workloads where your evals still prefer Claude Fable / Opus max effort, or where you need a provider feature Grok does not expose yet. A gateway that already speaks OpenAI Chat Completions makes that A/B cheap: same client, different model string.

Call it through an OpenAI-compatible gateway

If you already use OpenFluxHub (or any OpenAI-compatible relay), point the same client at grok-4.5 once the model is enabled on your account. Check live availability and gateway pricing on /pricing.

curl https://openfluxhub.com/v1/chat/completions \
  -H "Authorization: Bearer $OPENFLUXHUB_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "grok-4.5",
    "messages": [
      {"role": "user", "content": "Find the bug, fix it, then explain: function median(a){a.sort();return a[a.length/2]}"}
    ]
  }'

For SpaceXAI's native Responses API, the model id is the same grok-4.5. Prefer the gateway path when you want one key across Claude, DeepSeek, Grok, and the rest of your catalog.

Practical rollout checklist

1. Pin the model id (grok-4.5) in config — do not silently float to an alias in production until you trust alias policy.
2. Turn on prompt-cache routing keys for multi-turn agent sessions.
3. Start reasoning at high, then dial down on easy classification / rewrite traffic.
4. Cap max output and tool-loop depth; token efficiency helps only if you do not let agents ramble.
5. Compare cost and quality against your current Opus / Fable default on the same 20–50 task suite before a full cutover.
6. Watch region availability if you have EU-only keys — SpaceXAI noted EU API console access lagging the US launch.

Bottom line

Grok 4.5 is the clearest "Opus-class capability at workhorse pricing" bid SpaceXAI has made public so far: 500K context, coding- and agent-first tooling, $2/$6 list pricing, and immediate distribution through Grok Build and Cursor. The marketing claims are strong; your evals decide whether it becomes the new default. If you already route models through one OpenAI-compatible key, adding grok-4.5 is mostly a config change — start on /pricing, run a side-by-side, and keep the previous default as a fallback until the numbers settle.