grok-4.3 edges gpt-5.4 in a narrow, format-first fight

The scoreline says it all: grok-4.3 wins 36.0 to 34.0, and this was not a blowout. It was a precision contest, and grok-4.3 simply made fewer avoidable mistakes where the prompt’s constraints mattered most.

The split is clean. gpt-5.4 took python-redact-logs by being more robust on regex boundaries and invalid IPv4 handling — the better engineering answer, full stop. But grok-4.3 answered back on status-update-delay and meeting-notes-summary, and those wins were about compliance, not style points: calmer workplace tone, tighter formatting, and exact adherence to the requested output structure.

The messy-orders-to-json task was a wash, a tie in both correctness and quality. That matters because it removes any fantasy that one model was consistently cleaner across structured-output work; they were identical there.

Final call: grok-4.3 wins this matchup, but only because it was more reliable on instruction discipline. gpt-5.4 had the stronger technical edge on log redaction, yet it lost the head-to-head by being a little looser where format and tone were non-negotiable.

How they were tested

We ran 4 fresh text tasks, generated on the fly for this matchup so neither model could prepare in advance, and had gpt-5.4-mini score each one. grok-4.3 scored 36.0 to gpt-5.4's 34.0.

1. python-redact-logs

Write a Python 3 function redact_log(line: str) -> str for an internal support tool. It must replace any email address with [EMAIL] and any IPv4 address with [IP], while leaving timestamps, ports, and version numbers unchanged. Examples: "user meera.cho@northbay.dev failed from 10.24.8.19 at 2025-04-12 09:14" -> "user [EMAIL] failed from [IP] at 2025-04-12 09:14"; "proxy 10.24.8.19:443 upgraded to v2.10.3" -> "proxy [IP]:443 upgraded to v2.10.3". Return code only.

Winner: gpt-5.4 — Both solutions satisfy the core requirement, but B is more robust: it uses a stricter IPv4 pattern that avoids invalid octets and better boundary handling for emails and IPs. A is simpler and likely sufficient for the examples, but it can match invalid IP-like strings more loosely.

2. status-update-delay

Draft a workplace status update for the VP of Operations. Context: the Phoenix Yard scanner rollout is slipping by 6 days because 14 of 60 devices arrived with dead batteries and the vendor cannot replace them before Tuesday. Field training in Mesa already happened, so we want to preserve confidence and propose a revised plan: start with the 46 working units on Wednesday, finish the remaining devices the following Monday, and keep the June 28 inventory freeze unchanged. Tone: calm, accountable, no blame. Length: 120-150 words.

Winner: grok-4.3 — A is slightly better aligned with the requested workplace status-update tone: calm, accountable, and concise, with no extra framing. B is solid but includes a conversational lead-in and a bit more promotional language, making it slightly less direct.

3. meeting-notes-summary

Read these meeting notes and produce: (1) a 2-sentence summary, and (2) a JSON object with keys decision, owner, due_date, and risks. Notes: - Checkout latency spiked after the Cart service 3.14 deploy on Thursday. - Nia found the new coupon lookup calls promo-cache twice for guest users. - Reverting only that flag dropped p95 from 1.9s to 1.1s. - Marco wants a full rollback, but Priya said the release also contains the tax rounding fix requested by Finance. - Decision: keep 3.14 in place, disable guest coupon lookup, and patch forward next sprint. - Owner: Nia to open a hotfix PR today; Omar to monitor tonight's traffic. - Risk: guest checkout won't show promotional savings until patch lands. - Target date for patch: 2025-07-09.

Winner: grok-4.3 — A follows the requested format more closely by providing exactly a 2-sentence summary and a JSON object with the required keys. B is clear, but its owner field is structured as an object rather than the single value implied by the prompt, making it less compliant.

4. messy-orders-to-json

Convert the messy order lines below into valid JSON as an array of objects sorted by order_id ascending. Schema for each object: order_id (string), customer (string), items (array of strings), priority ("low"|"normal"|"high"), rush (boolean), total_usd (number with 2 decimals). Rules: trim spaces, normalize customer names to title case, split items on |, and interpret rush values Y/yes/true as true and N/no/false as false. Data: ORD-9081 ; customer= lENA ortiz ; items=label rolls|thermal head ; priority=high ; rush=Y ; total=$184.5 ORD-9077;customer=omar dunn;items= usb-c dock | hdmi cable |ethernet adapter;priority=normal;rush=no;total=249 ORD-9080 ; customer = PRIYA sen ; items = nitrile gloves | face shields ; priority = low ; rush = true ; total = $73.00

Winner: Tie — Both outputs are valid JSON, correctly sorted by order_id, and accurately apply the normalization and parsing rules. They are identical in content and quality.

See every prompt and the full side-by-side outputs in the interactive Head-to-Head.