Loading page
OpenAI
GPT-4 Turbo
GPT-4 Turbo is a chat model from OpenAI in the GPT-4 family, offering a 128K token context window and vision input alongside function calling and JSON output modes.
Input from
$5.00 / 1M tokens
across 1 provider
API Pricing
Cheapest on OpenRouter — 33% below avg| Provider | Input / 1M | Output / 1M |
|---|---|---|
| $5.00 | $15.00 | |
| $10.00 | $30.00 |
Prices updated daily. Last check: Sep 5, 2026
GPT-4 Turbo pricing by provider
Input, output, and batch rates, plus alternatives, for one provider at a time.
Performance & Benchmarks
Source: Artificial Analysis →Intelligence
2.3 / 100
Coding
21.5 / 100
Reasoning & Knowledge
- MMLU-Pro69.4%
- Humanity's Last Exam3.1%
Coding
- LiveCodeBench29.1%
Math
- AIME15.0%
- MATH-50073.7%
Benchmarks measured Sep 2026. Scores are independent evaluations, not vendor-reported.
Model Details
General
- Creator
- OpenAI
- Modalities
- Text
Capabilities
- Tool Calling
- No
- Open Source
- No
Strengths & Limitations
Strengths
- 128K token context window, far larger than the 8K/32K windows of earlier GPT-4 variants
- Accepts image inputs in addition to text for document, screenshot, and chart interpretation
- Function/tool calling with support for multiple parallel tool calls in a single turn
- JSON mode constrains responses to valid JSON, simplifying structured extraction pipelines
- Seed parameter allows more reproducible sampling for testing and regression suites
- Available under dated snapshot IDs, so production behavior can be pinned rather than tracking a moving alias
- Widely supported across SDKs, agent frameworks, and evaluation tooling due to long availability
Limitations
- Superseded within OpenAI's lineup by the GPT-4o and o-series generations, so it is not where new capabilities land
- Maximum output length is much smaller than its 128K input window, limiting very long generations
- Proprietary — served only through APIs, with no downloadable weights
- Fixed training knowledge cutoff means recent events require retrieval or tool augmentation
- No verified throughput or time-to-first-token measurements in our benchmark data for this entry
- Not a reasoning-tier model; math- and proof-heavy tasks may favor OpenAI's o-series or other reasoning models
Key Features
•128K token context window
•Vision input (image understanding alongside text)
•Function and tool calling, including parallel tool calls
•JSON mode for schema-constrained structured output
•Seed parameter for more reproducible outputs
•Streaming responses via the Chat Completions API
•Dated model snapshots for version pinning
•Available through OpenAI's API and Azure OpenAI Service
About GPT-4 Turbo
GPT-4 Turbo is a chat-oriented large language model from OpenAI, positioned within the GPT-4 family as the higher-throughput, longer-context successor to the original GPT-4 releases. It was the model line OpenAI pointed developers toward for general-purpose GPT-4-class work before the GPT-4o generation took over that role, and it remains available through the OpenAI API under dated snapshot identifiers (for example, gpt-4-turbo-2024-04-09) as well as the rolling gpt-4-turbo alias.
The model supports a 128K token context window, a substantial increase over the 8K and 32K windows of the earlier GPT-4 variants, which makes it usable for long-document analysis and multi-turn agent transcripts without aggressive truncation. GPT-4 Turbo accepts image inputs in addition to text, and the API exposes function/tool calling, parallel tool calls, JSON mode for schema-constrained output, and reproducible sampling via the seed parameter. Output is capped well below the input window, so it is better suited to reading large inputs than to producing very long single responses.
In practice GPT-4 Turbo is used for retrieval-augmented question answering over long sources, code assistance, structured data extraction, and tool-driven agent loops. Because OpenAI has since shipped newer GPT-4o and o-series models, GPT-4 Turbo is most often encountered today in existing production stacks that were built against it and in comparisons where teams are weighing whether to migrate. We do not currently have verified throughput or latency measurements for this model in our benchmark data, so compare provider-reported figures directly.
Common Use Cases
GPT-4 Turbo suits general-purpose GPT-4-class workloads where a large input window matters more than the newest capability set: summarizing and querying long contracts, filings, or research papers; retrieval-augmented chat over sizeable document sets; code explanation, refactoring suggestions, and review comments; and structured extraction where JSON mode plus function calling turns free text or screenshots into typed records. Vision input extends it to reading scanned pages, UI screenshots, dashboards, and diagrams. It is a reasonable fit for tool-driven agents that need multiple parallel tool calls per turn, and its dated snapshots make it workable for regulated or heavily tested deployments that need pinned behavior. For high-volume classification and cheap bulk labeling, smaller OpenAI tiers are usually the better economic choice, and for math- or science-heavy step-by-step reasoning, a dedicated reasoning model is the more direct option.
Frequently Asked Questions
How much does GPT-4 Turbo cost?
Pricing varies by provider, by pricing type (on-demand input and output token rates, batch discounts, cached input rates), and by contract. Input and output tokens are billed at different rates, and image inputs are billed based on resolution and detail settings. See the pricing table on this page for current per-provider figures.
What is GPT-4 Turbo best used for?
Long-context work: document question answering, retrieval-augmented chat, code assistance, and structured extraction from text or images. Its 128K context window plus JSON mode and parallel tool calling make it a practical fit for agent loops and data-extraction pipelines.
How does GPT-4 Turbo differ from the original GPT-4?
The main differences are a 128K context window instead of 8K/32K, vision input, and API features such as JSON mode, parallel tool calls, and the seed parameter for reproducibility. It was also positioned as the higher-throughput option in the GPT-4 family.
Should I use GPT-4 Turbo or a newer OpenAI model?
OpenAI has released GPT-4o and o-series models since GPT-4 Turbo, and those generations are where newer capabilities appear. GPT-4 Turbo is most often kept in place for existing integrations that are already validated against a specific dated snapshot. If you are starting fresh, benchmark a newer model on your own evals before committing.
Can GPT-4 Turbo process images?
Yes. GPT-4 Turbo accepts image inputs alongside text, which covers screenshots, scanned documents, charts, and diagrams. Image tokens are billed separately from text tokens and depend on the image size and detail level requested.
Does the 128K context mean it can write 128K tokens of output?
No. The 128K figure is the combined context window, and the per-request output limit is considerably smaller. Plan long-form generation as multiple calls rather than a single response.