Grok 4.6 vs Grok 4.5: We Ran 80 API Calls on Launch Day — Same 40/40, 1.38× the Bill

Grok 4.6 vs Grok 4.5: We Ran 80 API Calls on Launch Day — Same 40/40, 1.38× the Bill

Grok 4.6 cost us $0.2992 to finish 40 graded tasks. Grok 4.5 cost $0.2174 for the same 40, and returned the same answers.

We measured it on 12 August 2026, hours after SpaceXAI (formerly xAI) shipped the model. Same rate card: $2 per million input tokens, $6 output. Same score: 40/40 against 40/40. The gap is 1.90× the median output tokens on 4.6, which lands as a 38% larger invoice.

What SpaceXAI shipped on 12 August

SpaceXAI released Grok 4.6 on 12 August 2026 at $2 per million input tokens and $6 output, the same rate card as Grok 4.5. It went out at 08:56 PT per 9to5Mac, and the official announcement (retrieved 2026-08-12) leads on agentic coding while publishing no cost-of-operation, latency or token-consumption figure at all.

1. Same-day availability as grok-4.6 on the SpaceXAI API and x-ai/grok-4.6 on OpenRouter.
2. A 500K-token context window, unchanged from Grok 4.5.
3. $2/M input, $6/M output, plus a fast variant at twice the rate.
4. An Artificial Analysis Intelligence Index score of 61, framed by the vendor as matching GPT-5.6 Sol.
5. One qualitative 4.5 comparison: stronger first passes on visual and interactive projects.

Cursor's launch-day write-up reads like third-party validation and is not: SpaceX agreed to buy Cursor's maker, Anysphere, for $60B in stock on 16 June 2026.

Should you switch to Grok 4.6?

Stay on 4.5 for routine structured work: our 80 calls returned identical answers for 1.38× the money. The rate card is byte-identical on both OpenRouter model pages, so no pricing page can warn you. Token counts can.

Rows marked "ours" are ours, measured on launch day; index and cache rows are Artificial Analysis's, retrieved 2026-08-12.

Identical rate card, 1.90× the tokens, therefore roughly 1.38× the invoice for the same answers. That is the grok 4.6 vs grok 4.5 question for most production traffic: same price per token, a different bill.

What 80 API calls on launch day actually showed

Across 80 launch-day calls at temperature zero, both models scored 40/40 while 4.6 spent 1.90× the median output tokens.

We sent each prompt twice, once to x-ai/grok-4.6 and once to x-ai/grok-4.5, measured through OpenRouter at temperature: 0. Round 1 (24 calls) was easy: the JSON-schema tasks we graded, a pricing calculation, a Python bug fix, a constrained writing task. Round 2 (24 calls) got harder after round 1 saturated: a scheduling puzzle, a unit-conversion trap, a 402-line needle retrieval with a REVOKED decoy, and a spec task where retry_on must contain 429 and 503 but not 500. Round 3 (32 calls) is the control below. Every grader is deterministic; nothing is LLM-judged. Scripts grok_bench.py, grok_bench_hard.py, grok_bench_effort.py; raw output in benchmark-raw.json, benchmark-hard-raw.json, benchmark-effort-raw.json. Total spend, $0.52.

for model in ("x-ai/grok-4.6", "x-ai/grok-4.5"):
    r = httpx.post("https://openrouter.ai/api/v1/chat/completions",
        headers={"Authorization": f"Bearer {KEY}"},
        json={"model": model, "temperature": 0,
              "messages": [{"role": "user", "content": PROMPT}]}).json()
    u = r["usage"]
    print(model, u["completion_tokens"],
          u["completion_tokens_details"]["reasoning_tokens"])

The pre-launch consensus, an X post by @haider1 dated 11 August and copied across aggregator blogs, said 4.6 would keep 4.5's speed and token efficiency. SpaceXAI never claimed that; its announcement makes no token claim at all. Unite.AI noted on launch day that "no independent evaluation has yet confirmed" the model's claims, so here is one. Grok 4.6 used 409 median output tokens to Grok 4.5's 215 on identical prompts at temperature zero, 365 median reasoning tokens against 200, and 27,565 total against 13,929. Whatever 4.6 gains, it pays for with 1.90× the median output tokens.

The tail is where it hurts

Same prompt, temperature: 0, three trials of unit_trap in the default-effort rounds:

A 6.6× spread on 4.6 against 1.8× on 4.5, on byte-identical inputs. Sizing a timeout or a per-request token budget, that tail matters more than the median, and it argues for when the newer model is the wrong default.

The one task that ran the other way

On constraint_schedule, 4.6 was faster at the median in the default-effort rounds: 26.38 s against 34.21 s, on fewer output tokens too (1,644 against 1,710). Reporting only the numbers that fit a thesis is what readers and Google discount.

Is it the model, or is it the default?

Pinning reasoning_effort to the same value on both models does not close the gap; it widens it, from 1.51× to 2.91×. docs.x.ai (retrieved 2026-08-12) lists four levels (low, medium, high, xhigh) and rounds 1 and 2 never set one, so the gap could have been a shipped default. Round 3 pinned it explicitly across 32 calls.

Had the gap collapsed at matched effort, the honest headline would have been "it's just a default." It did not.

How do you cut Grok 4.6's token bill?

Set reasoning_effort to low and 4.6's median output drops from 438 to 222 tokens, with accuracy unchanged at 8/8. Same four tasks either way: the default figure pools twelve calls from the first two rounds, the low figure eight from the control.

{
  "model": "x-ai/grok-4.6",
  "messages": [{"role": "user", "content": "..."}],
  "temperature": 0,
  "reasoning": {"effort": "low"}
}

That is a 49% reduction, worth roughly $1.30 per thousand calls of this shape at $6 per million output tokens. One request parameter roughly halves Grok 4.6's output bill on routine structured work, and gave up nothing we could measure. Four tasks is a signal, not a policy: test it on your own mix first.

What other people's meters show

Artificial Analysis, on its own evaluation suite, billed $1,068.47 to score Grok 4.6 against $579.21 for Grok 4.5. Their numbers, not ours, from their Grok 4.6 and Grok 4.5 model pages at artificialanalysis.ai, retrieved 2026-08-12.

Their 1.84× and our 1.38× disagree, and the reason is informative: their task mix is heavier, and their total folds in input tokens where ours isolates output. Two meters, same direction.

The cache row was flagged first by HN user pzo in the launch thread (comment 49275740) and confirms on both pages: up 67%, and it bites hardest on the long-running agent workloads 4.6 targets, where cache reuse is the whole point.

Is Grok 4.6 better than Claude for coding?

On correctness it matches: Grok 4.6, Claude Sonnet 5 and Claude Opus 4.8 each passed 10 of 10 executed coding tests. That is the headline, and for Grok a real gain.

We stopped grading text for this one. Five tasks with hidden unit tests, two trials each, 30 calls: semver matching including the ^0.x case, an LRU cache with explicit-clock TTL, interval merging with touching bounds, a duration parser that must reject 1m30h and 1.5h, and grapheme-safe truncation that must not orphan a combining mark or split an emoji. Each answer runs in a subprocess, passing only if every assertion holds. Script grok_vs_claude_coding.py, raw in benchmark-coding-raw.json.

p = subprocess.run([sys.executable, path], capture_output=True, timeout=20)
ok = p.returncode == 0 and "ALLPASS" in p.stdout   # the model never sees the tests

The correctness tie is the news. The bill is the catch: Grok 4.6 ran 4.0× slower than Sonnet 5 and 5.4× slower than Opus 4.8, on 4.0× and 4.9× the median output tokens. That token appetite eats the price advantage whole. Grok 4.6 cost more than Claude Opus 4.8 on these five tasks despite its output tokens being 4.2× cheaper, because it emitted 18,345 output tokens against Opus 4.8's 3,630. Against Sonnet 5 it cost 1.96× as much, and Sonnet's per-token output rate is the higher of the two. A cheaper rate card is not a cheaper model, which is the same lesson the 4.6-against-4.5 numbers teach one section up.

One judgement, flagged as judgement rather than measurement: on media and content-automation pipelines we have historically reached for Grok over Claude, mainly for its X-native ingestion and looser handling of marketing copy. We have no benchmark for that and are not presenting one. On the coding work we can grade deterministically, the three tie on correctness and Grok pays four to five times the tokens.

What we did not test

Our tasks saturated at 40/40, so this measures cost on routine work, not capability on the agentic jobs SpaceXAI tuned for. Five limits:

1. The accuracy tie is a ceiling effect, not proof 4.6 is no smarter. Our tasks ran out of difficulty before the models did, and they do not probe frontier long-horizon coding.
2. Two to three trials per task. Enough for a direction and a variance story, not a confidence interval.
3. Measured through OpenRouter, not SpaceXAI's own endpoint. Routing adds latency that is not the model's, which is why the provider-independent token counts carry the argument.
4. One task ran against the thesis, constraint_schedule, where 4.6 was faster at the median.
5. The Claude comparison saturated too. Five coding tasks, all three models 10/10, so it separates cost and latency but says nothing about which model is stronger at the ceiling.

We also did not test what SpaceXAI actually claims: stronger first passes on visual and interactive projects. No deterministic grader exists for that, so we are not contesting it. We hold no commercial relationship with SpaceXAI and paid for every call ourselves.

Who should upgrade, and who should stay on 4.5?

Upgrade if your traffic is long-horizon agent work; stay on 4.5 if it is short structured calls at volume. On the index axis every other launch-day article uses, 4.6 wins at an identical rate card. On measured cost per correct answer, grok 4.6 vs grok 4.5 flips the other way.

Grok 4.5 is still the cheaper way to get the identical answer on routine structured work. That is not the same as 4.6 being worse: it buys its gains by thinking longer, and on everyday production calls that trade is a cost increase with no accuracy return we could detect. One more argument for pinning a model version in an agent stack instead of tracking latest.

Three scripts, one OpenRouter key and under a dollar of credit is the whole cost of checking whether your traffic looks like ours.

Frequently Asked Questions

Is there a Grok 5 or Grok 5.6?

Neither exists. As of 12 August 2026, Grok 4.6 is the newest shipped model. Grok 4.7 was signalled for late August or early September and had not shipped when we wrote this; anything past it is targeted for end of 2026. The "5.6" is almost certainly GPT-5.6 Sol, an OpenAI model Grok 4.6 is benchmarked against.

Does Grok 4.6 cost more than Grok 4.5?

Same rate card, $2/M input and $6/M output on both. Our 80 measured calls returned identical answers for 1.38× the total cost, because 4.6 emits 1.90× the median output tokens. Cached input also rose from $0.30 to $0.50 per million.

Is Grok 4.6 slower than Grok 4.5?

At our median, 1.26× slower: 5.25 s against 4.17 s across 40 calls each. At our worst call, 2.27×: 81.61 s against 35.98 s. Artificial Analysis, measuring separately, reports time to first token at 32.30 s versus 8.68 s. Note that we measured through OpenRouter, so routing adds latency the model itself is not responsible for.

What is Grok 4.6's context window and how do I call it?

500K tokens, unchanged from Grok 4.5. The slug is x-ai/grok-4.6 on OpenRouter and grok-4.6 on the SpaceXAI API, with a fast variant at twice the standard rate. Set reasoning_effort explicitly rather than inheriting it; the default is high, and on our four control tasks dropping it to low halved the output tokens without costing a single correct answer.

Should I stay on Grok 4.5?

For high-volume structured work, yes: it returned the same answers for less money across all 80 calls. For long-horizon agentic coding, the workload SpaceXAI tuned 4.6 for, our tasks do not settle it and we did not test it. Measure your own mix.