Migrating from gpt-5.2 to gpt-5.4

What changes when you move from gpt-5.2 to gpt-5.4: input price +42.9%, output price +7.1%, context window 400,000 → 1,050,000 tokens.

From: gpt-5.2
Status: Active
Shutdown: none announced
Time left: —

To: gpt-5.4
Status: Active
Shutdown: none announced
Time left: —

Prices

USD per 1M tokens, standard tier

Type	gpt-5.2	gpt-5.4	Change
Input	$1.75first seen October 1, 2026 · source	$2.50first seen October 1, 2026 · source	+42.9%
Cached input	$0.175first seen October 1, 2026 · source	$0.25first seen October 1, 2026 · source	+42.9%
Output	$14.00first seen October 1, 2026 · source	$15.00first seen October 1, 2026 · source	+7.1%

At 1M input and 250K output tokens a month: $5.25 on gpt-5.2, $6.25 on gpt-5.4 (+19%).

Other tiers, prompt-size bands and modalities are on each model's page.

Limits

Tokens

Limit	gpt-5.2	gpt-5.4	Change
Context window	400,000 tokensin place when tracking began (October 6, 2026) · source	1,050,000 tokensin place when tracking began (October 7, 2026) · source	+162.5%
Max output tokens	128,000 tokensin place when tracking began (October 6, 2026) · source	128,000 tokensin place when tracking began (October 7, 2026) · source	0%

Benchmarks

Same benchmark and source on both sides; settings named where they differ

Benchmark	gpt-5.2	gpt-5.4	Change
Arena TextArena, high effort. Ranges don't overlap. Other settings 1435 ±3 → 1465 ±4	1437 ±4	1475 ±4	+38
Arena Text: CodingArena, high effort. Ranges don't overlap. Other settings 1482 ±5 → 1512 ±6	1489 ±6	1520 ±6	+31
Arena WebDevArena, high effort, codex harness (vs no setting). Ranges don't overlap. Other settings not rated → 1444 ±18 medium effort, codex harness 1417 ±23 → 1402 ±10	1417 ±23	1463 ±19	+46
Arena VisionArena, high effort. Ranges don't overlap. Other settings 1229 ±6 → 1279 ±7	1243 ±6	1285 ±6	+42
Arena SearchArena. Ranges don't overlap.	1161 ±6	1196 ±6	+35
Arena DocumentArena. Ranges don't overlap.	1425 ±6	1470 ±6	+45
FrontierMath Tier 4Epoch AI, xhigh effort.	31.7%	49.0%	+17.3 pts
FrontierMath Tiers 1-3Epoch AI, xhigh effort.	67.4%	78.6%	+11.2 pts
GPQA DiamondEpoch AI, xhigh effort. Other settings 88.2% → 89.9% high effort 87.9% → 88.9% medium effort 82.7% → 84.8% low effort 73.2% → 74.7% no reasoning	91.4%	93.3%	+1.9 pts
OTIS Mock AIME 2024-2025Epoch AI, high effort. Other settings 93.9% → 95.6% medium effort 96.1% → 95.3% xhigh effort 78.9% → 84.4% low effort 62.2% → 57.8% no reasoning	96.1%	97.8%	+1.7 pts
SimpleQA VerifiedEpoch AI, xhigh effort.	37.1%	45.1%	+8.0 pts
SWE-bench VerifiedEpoch AI, high effort.	73.8%	76.9%	+3.1 pts

A score from one source is never compared with another source's, and company-reported scores are compared only within the same company; settings named where they differ. Data: Arena (arena.ai), CC BY 4.0; Epoch AI, Capabilities & benchmarking (epoch.ai/benchmarks), CC BY.

Features and parameters

Compare models

Same on both

Structured outputs: supported on both.gpt-5.2: in place when tracking began (October 6, 2026) · source · gpt-5.4: in place when tracking began (October 7, 2026) · source
Function calling: supported on both.gpt-5.2: in place when tracking began (October 6, 2026) · source · gpt-5.4: in place when tracking began (October 7, 2026) · source
Prompt caching: supported on both.gpt-5.2: in place when tracking began (October 6, 2026) · source · gpt-5.4: in place when tracking began (October 7, 2026) · source
Batch API: supported on both.gpt-5.2: in place when tracking began (October 6, 2026) · source · gpt-5.4: in place when tracking began (October 7, 2026) · source
Image input: supported on both.gpt-5.2: in place when tracking began (October 6, 2026) · source · gpt-5.4: in place when tracking began (October 7, 2026) · source
Reasoning / thinking: supported on both.gpt-5.2: in place when tracking began (October 6, 2026) · source · gpt-5.4: in place when tracking began (October 7, 2026) · source
Fine-tuning: not supported on both.gpt-5.2: in place when tracking began (October 6, 2026) · source · gpt-5.4: in place when tracking began (October 7, 2026) · source
Built-in web search: supported on both.gpt-5.2: in place when tracking began (October 6, 2026) · source · gpt-5.4: in place when tracking began (October 7, 2026) · source

Not published for gpt-5.2: PDF input, temperature parameter, top_p parameter, logprobs.

Not published for gpt-5.4: PDF input, temperature parameter, top_p parameter, logprobs.

Aliases

All aliases

No aliases point at either model.

On cloud platforms

All platform dates

Model	Platform	Platform ID	Shutdown	Time left
gpt-5.2	Microsoft Foundryversion 2025-12-11	gpt-5.2	Jun 8, 2027Active	236 days
gpt-5.4	Amazon Bedrock	openai.gpt-5.4	not before Jun 1, 2027Active
gpt-5.4	Microsoft Foundryversion 2026-03-05	gpt-5.4	Sep 2, 2027Active	322 days

Read daily from each platform's own pages. The status above uses each company's own date. Time left counts to 00:00 UTC on the shutdown date, which is the evening before in ET.

Checklist

Change the model string from gpt-5.2 to gpt-5.4.
Re-check costs: gpt-5.4 costs more than gpt-5.2 (input +42.9%, output +7.1%).

Compared from AIWire's published data when this page loaded. Each fact links to the page it came from. Model pages: gpt-5.2, gpt-5.4.