OpenAI models and capabilities

OpenAI’s model line in 2026 has three families on the platform API: the GPT generative line (GPT-5 / GPT-5-mini / GPT-5-nano + GPT-4o / GPT-4o-mini as workhorses, GPT-4 Turbo / GPT-3.5 as legacy), the o-series reasoning line (o3 / o3-mini / o4-mini / o1 / o1-pro — extended thinking by default; higher latency, higher quality on hard problems), and the specialty models (embeddings, Whisper, TTS, DALL-E 3 / gpt-image-1 for vision generation). The model-selection question is fundamentally tier (frontier vs cheap), reasoning depth (use the o-series only when you need it; reasoning tokens are billed and slow), and modality (Realtime for voice, gpt-image-1 for images, Whisper for transcription). The biggest 2025 change was GPT-5 taking over as default frontier model; the biggest API-surface change was the Responses API becoming the primary endpoint with Chat Completions still supported but no longer the default in new docs.

See also

• openai-api-and-sdks — how to reference these IDs via the Responses + Chat Completions APIs
• prompt-caching-openai — automatic caching mechanics + per-model rate
• tool-use-and-function-calling-openai — tool-use overhead
• hidden-tricks-and-gotchas-openai — undocumented quirks per model
• claude-models-and-capabilities — sister table for the Claude line
• gemini-models-and-capabilities — sister table for the Gemini line

1. GPT-5 family (current frontier, 2025-2026 release)

Per OpenAI announcements + platform.openai.com/docs/models as cross-referenced via developers.openai.com. GPT-5 launched in 2025 as the successor to GPT-4o; the family follows a similar three-tier shape.

Feature	GPT-5	GPT-5-mini	GPT-5-nano
API ID (alias)	gpt-5	gpt-5-mini	gpt-5-nano
Pinned snapshot	gpt-5-2025-XX-XX (date varies)	gpt-5-mini-2025-XX-XX	gpt-5-nano-2025-XX-XX
Approximate input price (USD / MTok)	~$2.50 - $5	~$0.25 - $0.50	~$0.05 - $0.10
Approximate output price (USD / MTok)	~$10 - $15	~$1 - $2	~$0.40 - $0.80
Context window	128k - 256k tokens (observed 200k typical)	128k+	128k+
Modalities	text + image (multimodal)	text + image	text
Vision support	Yes	Yes	Limited
Audio in/out	via Realtime (gpt-realtime)	via Realtime	n/a
Tokenizer	o200k_base (cl100k_base predecessor)	o200k_base	o200k_base
Use case	Frontier coding, reasoning, complex agentic	High-throughput production default	Cheap classification / extraction

Status note: pricing for GPT-5 has shifted multiple times during 2025-2026. The ranges above are observed; verify current rates on the live pricing page before committing budget. The dateless aliases (gpt-5, gpt-5-mini, gpt-5-nano) resolve to the latest pinned snapshot — pin to the dated ID for stability.

2. GPT-4o family (still production-ready)

The GPT-4o (“omni”) line was the dominant production default from 2024 into 2025 — still fully supported through 2026 and many teams remain on it for cost / stability reasons.

Feature	GPT-4o	GPT-4o-mini	GPT-4o-realtime
API ID	gpt-4o	gpt-4o-mini	gpt-4o-realtime-preview
Pinned snapshot	gpt-4o-2024-11-20 (and several others)	gpt-4o-mini-2024-07-18	gpt-4o-realtime-preview-2024-12-17
Input price (USD / MTok)	$2.50	$0.15	$5 text / $100 audio
Output price (USD / MTok)	$10.00	$0.60	$20 text / $200 audio
Cached input discount	50 % (e.g. $1.25 / MTok cached)	50 %	50 %
Context window	128k tokens	128k tokens	128k tokens
Max output	16,384 tokens (gpt-4o) / 16,384 (gpt-4o-mini)	same	same
Modalities (input)	text + image	text + image	text + image + audio
Modalities (output)	text + (audio via Realtime)	text	text + audio (real-time)
Knowledge cutoff	Oct 2023	Oct 2023	Oct 2023
Tokenizer	o200k_base	o200k_base	o200k_base

Pricing snapshot from platform.openai.com/docs/pricing reference:

• GPT-4o: $2.50 input / $10.00 output / $1.25 cached input — solid workhorse
• GPT-4o-mini: $0.15 input / $0.60 output / $0.075 cached input — best price/performance until GPT-5-mini

3. o-series reasoning models

Models that think before answering — internal chain-of-thought tokens (billed as output) precede the visible response. Generally higher latency, much higher quality on hard problems (competition math, code generation, scientific reasoning).

Feature	o3	o3-mini	o4-mini	o1	o1-pro
API ID	o3	o3-mini	o4-mini	o1	o1-pro
Status	Latest reasoning frontier	Cheap-tier reasoning	Hybrid reasoning + speed	Original reasoning model	Higher-compute o1 variant
Input price (USD / MTok)	~$15 (premium)	~$1.10	~$3	~$15	~$150
Output price (USD / MTok)	~$60	~$4.40	~$12	~$60	~$600
Context window	200k+	200k+	200k+	200k	200k
Max output (incl. reasoning tokens)	100k	100k	100k	32k (lower)	100k
reasoning.effort param	low / medium / high	same	same	low / medium / high	(preset high)
Tool support	Yes (function calling)	Yes	Yes	Limited (no streaming)	Limited
Vision	Yes	Limited	Yes	No	No
Stream output	Yes	Yes	Yes	No (response delivered atomic)	No
System messages	Treated as developer messages	same	same	same	same

Key parameter — reasoning.effort: controls how much the model “thinks” before answering. low is roughly 1k reasoning tokens; high can be 20k+. Billing includes reasoning tokens as output, so high-effort o3 calls can cost surprisingly much — a “simple” question on o3 with effort=high can produce $1+ of reasoning tokens before the visible answer.

When to use o-series:

• Hard math problems where GPT-5 / GPT-4o gets it wrong
• Code generation requiring multi-step reasoning
• Scientific Q&A demanding multi-hop logic
• Complex agent planning where one big think beats many small calls

When NOT to use o-series:

• Chat / dialog (high latency degrades UX)
• Streaming UX where partial output matters (o1 doesn’t stream)
• Simple extraction / classification (waste of reasoning tokens)
• Tool-heavy workflows (the model can reason → call tool → reason → ... but it’s expensive)

4. Legacy models

Feature	GPT-4 Turbo	GPT-4	GPT-3.5 Turbo
API ID	gpt-4-turbo	gpt-4	gpt-3.5-turbo
Status	Still callable, deprecating	Deprecated	Deprecated (still GA but discouraged)
Input / Output (USD / MTok)	$10 / $30	$30 / $60	$0.50 / $1.50
Context window	128k	8k - 32k	16k
Modalities	text + image (Turbo Vision)	text	text

Migrate GPT-4 / GPT-4 Turbo → GPT-4o / GPT-5; migrate GPT-3.5 → GPT-4o-mini / GPT-5-nano. Per OpenAI deprecations, legacy IDs continue to work but new builds should target current generation.

5. Embedding models

Per platform.openai.com/docs/guides/embeddings:

Model	Default dims	Max input tokens	Price (USD / MTok)	Notes
text-embedding-3-large	3,072	8,191	$0.13	Truncatable to smaller dims via dimensions param
text-embedding-3-small	1,536	8,191	$0.02	Default for most workloads; very cheap
text-embedding-ada-002 (legacy)	1,536	8,191	$0.10	Older; use only for back-compat

text-embedding-3-* supports the dimensions parameter for Matryoshka truncation — smaller dim → less storage / faster ANN at minor quality loss.

client.embeddings.create(
    model="text-embedding-3-large",
    input=["chunk 1", "chunk 2"],
    dimensions=1024,           # truncate to 1024 from the natural 3072
)

6. Audio models

Model	Purpose	Price	Notes
whisper-1	Speech-to-text (ASR)	$0.006 / minute	Translation also available
gpt-4o-transcribe	Transcription via GPT-4o	varies	Higher quality than Whisper for some languages
gpt-4o-mini-transcribe	Cheap transcription	varies
tts-1	Text-to-speech (standard)	$15 / 1M chars	6 built-in voices
tts-1-hd	Text-to-speech (HD quality)	$30 / 1M chars	Same voices, better quality
gpt-4o-tts	TTS via GPT-4o	varies	More expressive, longer-context

TTS voices: alloy, echo, fable, onyx, nova, shimmer — plus newer voices like ash, coral, sage, verse (varies by model). Choose per-call.

7. Image models

Model	Purpose	Pricing	Notes
dall-e-3	Image generation	$0.040 - $0.080 per image (size / quality)	Larger sizes + HD more expensive
gpt-image-1	Newer image gen (multi-input editing)	varies	Native image input + edit; more controllable
dall-e-2 (legacy)	Image generation	cheaper	Lower quality, used for variations

DALL-E 3 sizes: 1024x1024, 1024x1792, 1792x1024. Quality: standard or hd. Style: vivid or natural.

img = client.images.generate(
    model="dall-e-3",
    prompt="A cat astronaut on the moon.",
    size="1024x1024",
    quality="hd",
    style="vivid",
    n=1,
)
print(img.data[0].url)

8. Pricing details beyond per-token

Batch API

50 % discount on input + output via the Batch API. 24-hour SLA. See batch-and-fine-tuning-openai.

Automatic prompt caching

50 % discount on cached input tokens for any prompt ≥ 1,024 tokens. Automatic — no breakpoints. 5-10 minute cache window. See prompt-caching-openai.

Realtime API

Separate pricing — typically higher per-token for audio. See realtime-api-openai.

Fine-tuning

Operation	Cost
Training (gpt-4o-mini)	~$3 / 1M training tokens
Training (gpt-3.5-turbo)	~$8 / 1M training tokens
Inference (fine-tuned model)	Premium over base — varies per model
Reinforcement fine-tuning (RFT)	Preview pricing

See batch-and-fine-tuning-openai.

Hosted-tool surcharges (Responses API)

• file_search — small per-storage-day fee on vector store; per-search call cost
• web_search_preview — per-search fee
• computer_use_preview — premium per-token rate
• code_interpreter — per-container-hour rate

Exact rates change frequently; check current pricing page.

9. Programmatic model discovery

from openai import OpenAI
client = OpenAI()
for m in client.models.list():
    print(m.id, m.created, m.owned_by)

The list endpoint returns active models for your account. Note: model IDs in the list are pinned snapshots and aliases — gpt-5 and gpt-5-2025-XX-XX both appear when both are available. Filter accordingly.

10. Capability flags by model

10.1 Tool / function calling

Almost universal across current models. Notable exceptions:

• o1 — limited tool support (no streaming with tools)
• o1-pro — same
• Embedding / Audio / Image models — no tools (they’re not chat models)

10.2 Vision input

Model	Vision	Notes
GPT-5 family	✅	Multimodal native
GPT-4o family	✅	Multimodal native
o3 / o4-mini	✅	Yes (later additions)
o3-mini	Limited / no
o1 / o1-pro	❌	Text-only
GPT-4-turbo	✅	Vision GA
GPT-3.5	❌	Text-only

detail: "low" reduces image cost to a flat ~85 tokens; detail: "high" adds tiles. See vision-and-multimodal-openai.

10.3 Structured outputs (strict JSON schema)

Model	response_format: json_schema strict
GPT-5 family	✅
GPT-4o family	✅ (gpt-4o-2024-08-06 and later)
o3 / o4-mini	✅
o1	✅ (later snapshots)
Older GPT-4 / GPT-3.5	❌ — use response_format: json_object (loose JSON mode)

See structured-outputs-openai.

10.4 Prompt caching

Automatic for all current chat models when prompt ≥ 1,024 tokens. No opt-in. See prompt-caching-openai.

10.5 Parallel tool calls

Supported on all current chat models. Set parallel_tool_calls=False to force serial. o-series models may execute serially under the hood even with the flag on (reasoning step happens once, all tool calls picked together).

11. Tokenizer reference

Model family	Tokenizer	Use in tiktoken
GPT-5 family	o200k_base	tiktoken.encoding_for_model("gpt-5")
GPT-4o family	o200k_base	tiktoken.encoding_for_model("gpt-4o")
o-series	o200k_base	tiktoken.encoding_for_model("o1")
GPT-4 / GPT-4-turbo	cl100k_base	tiktoken.encoding_for_model("gpt-4")
GPT-3.5	cl100k_base	tiktoken.encoding_for_model("gpt-3.5-turbo")
text-embedding-3-*	cl100k_base	Same

The o200k_base switch matters for cost estimation — same English text often tokenizes ~10 % shorter on o200k_base than cl100k_base. Don’t reuse old token-count math when migrating from GPT-4-turbo to GPT-4o or GPT-5.

12. Endpoint types (cloud providers)

Surface	Hostname	Auth
OpenAI Platform	api.openai.com	API key
Azure OpenAI	<resource>.openai.azure.com	Azure AD / managed identity / API key
Workload Identity (cloud-native)	Same	Short-lived tokens from K8s / Azure / GCP service accounts

Azure OpenAI uses different model IDs (deployment names) but the same API shape via AzureOpenAI client. See openai-api-and-sdks.

13. Benchmark snapshot

Per OpenAI launch posts + developer blog. Treat as directional.

Benchmark	GPT-5	GPT-4o	o3	o3-mini	Notes
MMLU	mid 90s	low 90s	high 90s	mid 90s	Knowledge breadth
GPQA Diamond	mid 80s	high 50s	high 80s	high 70s	Graduate STEM (o-series shine)
AIME 2024 (math)	high 80s	low 20s	mid 90s	mid 80s	o-series specialty
SWE-bench Verified	strong	mid	strong	strong	Coding
HumanEval	saturated (~95+)	high 80s	mid 90s	low 90s	Saturated; rarely cited

Per OpenAI engineering posts 2025-Q4. The o-series numbers on AIME and GPQA are the headline — reasoning models materially outperform regular GPT-5/4o on hard math + science.

14. Versioning convention

OpenAI’s model IDs follow <base>[-<minor>]-<YYYY-MM-DD>:

• gpt-4o-2024-11-20 — pinned snapshot
• gpt-4o — alias resolving to the latest pinned snapshot
• gpt-4o-mini — alias
• gpt-4o-mini-2024-07-18 — pinned

For production, always pin to the dated ID. Aliases silently advance to new snapshots when OpenAI ships an update — behavior can change underneath you.

For Azure OpenAI, you bind a deployment name to a model version at deploy time; the deployment name is what you call in the API.

15. Further reading

• Models — canonical comparison
• Pricing — per-model + per-feature
• Model release notes — chronological history
• Reasoning models guide — o-series usage patterns
• Embeddings guide — embedding model details
• Tokenization (tiktoken) — official tokenizer library