GPT-4o fits teams still running production integrations on a pinned snapshot who value stability over new capability. For anything new, look elsewhere: OpenAI's own comparison put GPT-4.1 21.4 points ahead on SWE-bench Verified, and GPT-5.x has widened the lead since. Migrate agentic or coding workloads first.
GPT-4o is OpenAI's natively multimodal model, scoring 90.2% on HumanEval and processing text, image, audio, and video through one end-to-end network rather than stitching together separate vision and speech systems. It once ranked among the strongest general-purpose models available, before GPT-5.x overtook it on every current benchmark.
Where it sits
- $4.38/M$ per 1M tokensBlended price (3:1)Lower is better#39 / 60peer median $2.08/Mvendor price, checked by HokAI
- --tokens/sOutput speedHigher is better-- / 33peer median 90 tok/scited: Artificial Analysis
- 33.2%% solvedSWE-bench VerifiedHigher is better#27 / 27peer median 78%per source, see benchmark scores
- --% correctGPQA DiamondHigher is better-- / 41peer median 86.9%per source, see benchmark scores
Priced around the middle of the 60 GA models with a published price (rank 39), in the bottom third on SWE-bench Verified (rank 27 of 27), and one of 20 that document a zero-data-retention option. Ranked against GA models; this record is not GA.
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: OpenAI · Family: GPT-4o
Context window: 128,000 tokens · Max output: 16,384
Input modalities: text, image, audio, video, tool-calls · Output: text, audio, image, tool-calls
About GPT-4o
GPT-4o (the "o" stands for omni) is OpenAI's natively multimodal flagship model, released on May 13, 2024 during the company's Spring Updates livestream. Unlike earlier GPT-4 variants that added vision or voice onto a text-first model, GPT-4o was trained end-to-end across text, image, and audio in a single network, aimed at bringing GPT-4 Turbo-level reasoning to real-time, low-latency multimodal interaction at a fraction of the previous cost. OpenAI has never disclosed its parameter count or whether the architecture is dense or mixture-of-experts; independent estimates place it in the hundreds of billions, in line with the rest of the GPT-4 family. A smaller GPT-4o mini variant followed two months later for cost-sensitive workloads. On released benchmarks, GPT-4o scored 88.7% on MMLU, an improvement over GPT-4 Turbo's roughly 86.5% MMLU at the time, with equally strong HumanEval coding results. OpenAI never published a direct SWE-bench Verified score for GPT-4o, but when introducing GPT-4.1 in 2025 it stated that model improved on SWE-bench Verified by 21.4 points over GPT-4o, implying a GPT-4o baseline near 33%, a wide gap to the coding scores current frontier models post today. On the LMArena Chatbot Arena, the August 2024 snapshot reached an Elo near 1314; GPT-4o no longer appears on current leaderboards. GPT-4o ships with a 128,000-token context window. Maximum output was capped at 4,096 tokens at launch; a later snapshot raised that ceiling to 16,384 tokens, a change that matters for long-form generation, code output, and structured JSON responses. There is no separate extended-context tier, so 128K reads as mid-pack by current standards. As a natively multimodal model it accepts any mix of text, image, audio, and video as input, and its real-time voice mode runs through a dedicated Realtime API rather than the standard Chat Completions endpoint. GPT-4o is available through the OpenAI API and Azure OpenAI Service, with SDKs in Python, Node.js/TypeScript, Java, .NET, and Go. Fine-tuning reached general availability on a later dated snapshot. As a closed, proprietary model there are no downloadable weights or self-hosting option; access is strictly API-based. OpenAI published a dedicated GPT-4o System Card covering pretraining data filtering and post-training red-teaming, with particular attention to risks the real-time speech-to-speech capability introduces, including voice cloning and audio content moderation. OpenAI retired GPT-4o from ChatGPT on February 13, 2026, with Business, Enterprise, and Edu customers keeping access inside Custom GPTs only for a short window afterward. It remains reachable via the API but is officially deprecated, with OpenAI directing developers to GPT-4.1 for cost- and latency-sensitive workloads or GPT-5.x for frontier capability. Teams already running a pinned production snapshot can keep doing so at a known price; greenfield projects should default to GPT-4.1 or GPT-5.x instead, since GPT-4.1 already closes most of the SWE-bench Verified gap and GPT-5.x extends it further. GPT-4o's training data cutoff was originally October 2023, but a March 2025 update silently extended it to June 2024 without a model version change, so two requests to the same model string can reflect different effective knowledge cutoffs depending on when they were served. OpenAI's standard API data retention applies: inputs and outputs may be retained for up to 30 days for abuse monitoring unless an organization qualifies for zero-data-retention. OpenAI maintains SOC 2 Type II certification for the API platform and offers HIPAA business associate agreements and EU data residency options for eligible enterprise customers.
Pricing
GPT-4o costs $2.50 per 1M input tokens and $10.00 per 1M output tokens, the rate OpenAI has held since an October price cut from the original launch price of $5.00 input / $15.00 output. Cached input is $1.25 per 1M tokens. The Batch API gives a further 50% discount on both input and output for jobs returned within 24 hours.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.075 | $0.010 | $0.085 |
| Support reply | $0.0050 | $0.0030 | $0.0080 |
| One coding agent run | $0.500 | $0.200 | $0.700 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- Native Multimodality: Trained end-to-end on text, image, audio, and video together, instead of bolting vision or speech onto a text-first model.
- Real-Time Voice: Responds to spoken input in as little as 232 milliseconds (average 320ms) through OpenAI's separate Realtime API.
- 128K Context, 16,384-Token Output: A 128,000-token context window paired with a 16,384-token max output, four times GPT-4o's original launch cap.
- Function Calling With Vision: Lets the model inspect an image and decide which tool to call in the same turn, a capability added after launch.
- Fine-Tuning and Structured Outputs: A later snapshot supports general-availability fine-tuning and strict JSON-schema structured outputs, so responses can conform exactly to a defined schema.
Pros
- Handles voice, vision, and text natively, so teams skip stitching together separate speech-to-text, vision, and language services.
- Real-time voice response is fast enough for natural back-and-forth conversation, not just chat with a delay.
- 88.7% MMLU at release was a genuine step up over GPT-4 Turbo, a strong showing for a model of its era.
- The October price cut left it priced within reach of budget-conscious teams, even against newer competing models.
Cons
- Officially deprecated and retired from ChatGPT, with OpenAI steering all new development toward its newer model lineup instead.
- Inferred SWE-bench Verified near 33%, trailing current frontier models by a wide margin on agentic coding tasks.
- Smaller context window than most current frontier models, with no extended-context tier.
- Its effective knowledge cutoff moved once without any version bump, so pinning a model string is not a reliable way to know what it actually knows.
Benchmarks
- MMLU: 88.7% vendor-reported · 13 May 2024 — General-knowledge exam across 57 subjects, % correct.
- HumanEval: 90.2% vendor-reported · 13 May 2024 — Small programs that must pass hidden tests, % passing.
- LMArena Elo: 1314 vendor-reported · 13 May 2024 — Rating from blind human votes on which answer is better.
- SWE-bench Verified: 33.2% vendor-reported · 13 May 2024 — Real GitHub issues fixed end to end, % solved.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
How much do you pay for GPT-4o?
Input tokens run $2.50 per 1M and output tokens run $10.00 per 1M, unchanged since OpenAI's last price cut. Cached input tokens run $1.25 per 1M, and the Batch API cuts both input and output by a further 50% for asynchronous jobs returned within 24 hours. There's no subscription tier: every request is billed per token through the OpenAI API or Azure OpenAI Service.
Is GPT-4o free to use?
GPT-4o has no free tier at all; access is metered from the first request, with no standing free quota. New OpenAI platform accounts typically get a small one-time trial credit rather than ongoing free access to this model. Anyone wanting a genuinely free way to test similar multimodal capability should look at a model with an actual free API tier instead.
What should you use instead of GPT-4o?
For new work, GPT-5 is OpenAI's own direct upgrade path, with materially better reasoning and coding scores at competitive pricing. Claude Opus 4.8 and Gemini 3.1 Pro are worth considering too if the budget allows a premium model, both scoring well ahead of GPT-4o on current benchmarks. None of the three match GPT-4o's exact real-time voice latency, so voice-first products should test that path specifically before switching.
Is GPT-4o better than GPT-5?
GPT-5 beats GPT-4o decisively on reasoning and coding, posting 74.9% on SWE-bench Verified against GPT-4o's much lower inferred score, plus a larger context window and native reasoning modes GPT-4o never had. GPT-5's per-token API pricing also undercuts GPT-4o's. The one edge GPT-4o keeps is a dedicated low-latency Realtime voice API, which GPT-5 handles differently.
How long does it take to get going with GPT-4o?
Getting started with GPT-4o just needs an OpenAI API key: create one at platform.openai.com, then call the Chat Completions endpoint with the model pinned to a specific dated snapshot rather than the floating alias, so behavior doesn't shift under you. For voice, connect to the separate Realtime API over a websocket instead of Chat Completions. Existing OpenAI or Azure OpenAI credentials work immediately since GPT-4o needs no separate enrollment.
Top Alternatives
- GPT-5: Pick GPT-5 for any new coding or reasoning project: it scores far higher on SWE-bench Verified at a lower per-token price than GPT-4o.
- Claude Opus 4.8: Pick Claude Opus 4.8 for long-horizon coding and complex reasoning, where its far larger context window and current benchmarks leave GPT-4o's older scores behind.
- Gemini 3.1 Pro: Pick Gemini 3.1 Pro for the largest working context and native search grounding; pick GPT-4o only if you're maintaining an existing integration and want stability.