GPT-5.4

GPT-5.4 is a large language model developed by OpenAI and released on March 5, 2026.^[1] It is the fourth point release in the GPT-5 series, following GPT-5.3 (the Codex line and GPT-5.3 Instant) and preceding GPT-5.5. GPT-5.4 was the first time OpenAI shipped a single mainline model with native computer use, a 1 million token context window, and the coding capability that had previously been confined to its specialist Codex variants.^[1][10] OpenAI described it as a model that "brings together the best of our recent advances in reasoning, coding, and agentic workflows into a single frontier model" and called it "our most capable and efficient frontier model for professional work."^[1]

GPT-5.4 launched in two reasoning configurations at debut, GPT-5.4 Thinking and GPT-5.4 Pro, with two smaller siblings (GPT-5.4 mini and GPT-5.4 nano) following on March 17, 2026.^[1][2] A defensive cybersecurity variant, GPT-5.4-Cyber, was added later in April through OpenAI's Trusted Access for Cyber program.^[21][22] The model headlined a unified release across ChatGPT, the OpenAI API, and Codex surfaces, and it became the default "Thinking" tier in ChatGPT for paid plans on launch day.^[1][6]

OpenAI positioned GPT-5.4 as the first frontier model that combines coding, knowledge work, computer use, and long context in one general-purpose system.^[1] On OSWorld-Verified, a computer-use evaluation that measures how well models can drive desktop applications, GPT-5.4 scored 75.0%, exceeding the 72.4% average human baseline reported by OpenAI.^[1][12] On the GDPval professional knowledge work benchmark spanning 44 occupations, the model reached 83%.^[1] One AI news site reported that it briefly took the top Elo position on LM Arena the day after launch.^[27] Reception in the developer community was generally positive on capability, with criticism focused on price increases relative to predecessors and a time to first token of roughly three minutes in reasoning mode as measured by Artificial Analysis.^[8]

Why did OpenAI build GPT-5.4?

GPT-5.4 arrived about seven months after the original GPT-5 launch in August 2025. By early 2026 OpenAI had iterated rapidly: GPT-5.1 refined instruction following, GPT-5.2 introduced the three-tier Instant/Thinking/Pro structure on December 11, 2025, and the GPT-5.3 cycle in February and March 2026 produced GPT-5.3-Codex, GPT-5.3-Codex-Spark (with Cerebras), and GPT-5.3 Instant.^[15] By the time GPT-5.4 was being prepared, OpenAI's lineup had bifurcated into a conversational track (the Instant series) and a specialist coding track (the Codex series), each with separate price points and rollout schedules.^[15]

The positioning of GPT-5.4 reversed that trend. OpenAI described it as the model that "brings together the best of our recent advances in reasoning, coding, and agentic workflows into a single frontier model," folding the gains of GPT-5.3-Codex back into the mainline rather than maintaining a parallel Codex specialist.^[1] Brendan Foody, CEO of evaluation platform Mercor, said the model delivered "top performance while running faster and at a lower cost than competitive frontier models" on Mercor's APEX-Agents benchmark for law and finance.^[1][6]

The competitive context was a crowded frontier. Anthropic had released Claude Opus 4.6 on February 5, 2026, with a 1 million token context window in beta.^[31] Google's Gemini 3 family held a substantial share of long-context workloads through Vertex AI and offered cheaper input pricing. xAI, DeepSeek, and Meta had all shipped frontier or near-frontier models in the same window. OpenAI's pitch with GPT-5.4 was that the same model could now handle the long-context, computer-use, and agentic workflows that had previously required specialist tools or competing providers.

When was GPT-5.4 released?

GPT-5.4 launched on Thursday, March 5, 2026 across ChatGPT, the API, and Codex.^[1][6] OpenAI published the announcement "Introducing GPT-5.4" on its blog the same day, and GPT-5.4 Thinking began rolling out that day to ChatGPT Plus, Team, and Pro users, replacing GPT-5.2 Thinking, while Enterprise and Edu admins could enable early access.^[1] Free-tier users did not receive direct access to the base GPT-5.4 model at launch, although the smaller GPT-5.4 mini reached free users through the Thinking feature when it shipped on March 17, 2026.^[2][11]

In the API, GPT-5.4 was exposed under the model ID gpt-5.4 with a dated snapshot gpt-5.4-2026-03-05 for reproducible research.^[3] The Pro variant, which can spend longer in extended reasoning and is intended for high-stakes work, was exposed as gpt-5.4-pro.^[3] OpenAI also published an updated system card addendum tied to the existing GPT-5 system card line, classifying GPT-5.4 Thinking as High capability in the Cybersecurity domain under the Preparedness Framework.^[1][28]

The release was synchronized with several first-party surface updates. ChatGPT's web client added an upfront thinking plan preview that showed users what the model intended to do before it executed.^[1] Codex CLI and the IDE extensions for Visual Studio Code and JetBrains added GPT-5.4 as a selectable model.^[1] GitHub Copilot made the model generally available to Copilot Pro, Pro+, Business, and Enterprise users the same day.^[34] Microsoft made GPT-5.4 generally available in Microsoft Foundry on the same day.^[35]

In ChatGPT, GPT-5.2 Thinking stayed available to paid users under Legacy Models for three months, with retirement set for June 5, 2026.^[1] GPT-5.3-Codex remained available but was effectively superseded as a recommended option, with Codex documentation pointing developers toward GPT-5.4 for new projects.^[14]

What variants of GPT-5.4 are there?

GPT-5.4 launched with two reasoning configurations and added small models, computer-use, and a cyber-permissive variant in subsequent weeks.

GPT-5.4 Thinking

GPT-5.4 Thinking is the mainline reasoning model. It supports adjustable reasoning effort (none, low, medium, high, and xhigh), accepts text and image input, and produces text output.^[3] It is the version surfaced in ChatGPT's Thinking tier for paid users, in Codex CLI, and on most third-party platforms.^[1] OpenAI describes it as its "most capable and efficient frontier model for professional work," and Mercor's CEO singled out long-horizon deliverables such as slide decks, financial models, and legal analysis.^[1]

Thinking inherits the developer tools from earlier GPT-5 releases (the apply_patch tool for structured file diffs, local_shell for sandboxed shell execution, preambles for persistent agent instructions) and adds two new ones at launch: native computer use, which lets the model take screenshots, move a mouse, and type into desktop applications, and tool search, which dynamically loads tool definitions only when needed instead of front-loading them in every prompt.^[1][10][24]

GPT-5.4 Pro

GPT-5.4 Pro is the higher-end reasoning configuration. It supports medium (default), high, and xhigh reasoning effort and is available only through the Responses API. OpenAI positions Pro for tasks where additional inference compute reliably improves outcomes, such as legal due diligence, complex financial modeling, frontier scientific reasoning, and difficult coding tasks that require deeper deliberation.^[1] On Humanity's Last Exam Pro reaches 41.6%, second only to Gemini 3.1 Pro Preview at 44.7% according to Artificial Analysis tracking through May 2026.^[9]

Pro carries the same 1.05M-token context window as Thinking but a much steeper price tag: $30 per million input tokens and $180 per million output tokens, twelvefold the standard tier.^[9] The variant is available to ChatGPT Pro and Enterprise subscribers and through the API for all customers.

GPT-5.4 mini and nano

OpenAI added GPT-5.4 mini and GPT-5.4 nano on March 17, 2026, in a follow-up post titled "Introducing GPT-5.4 mini and nano."^[2] Mini retains computer use and tool search but trims the context window to 400,000 tokens.^[4] It is OpenAI's recommended option for high-volume agent and subagent workloads where the standard Thinking model is too expensive.^[2] In ChatGPT, mini is available to Free and Go users as the Thinking option, giving free-tier users their first access to a GPT-5.4 derivative.^[2][11]

Nano is the smaller of the two and is API-only. Computer use and tool search are not supported.^[5] It is targeted at high-volume classification, data extraction, ranking, and short-scope sub-agent work where speed and cost dominate the budget.^[5] Pricing for the small variants is substantially below the mainline tier.

GPT-5.4-Cyber

GPT-5.4-Cyber is a fine-tune of GPT-5.4 Thinking trained to support legitimate defensive cybersecurity work. OpenAI announced it on April 14, 2026, alongside an expanded Trusted Access for Cyber (TAC) program.^[21][22] The variant lowers the refusal boundary for tasks that the standard model would treat as borderline (vulnerability research, malware analysis, binary reverse engineering) and adds capabilities such as binary reverse engineering of compiled software without source code access.^[21] Access is restricted to identity-verified individuals and enterprise security teams, with optional Zero-Data-Retention waivers that allow OpenAI monitoring in exchange for higher permissiveness.^[21][22]

What are GPT-5.4's technical specifications?

The 1.05M-token context window was the largest in the GPT-5 series to that point (OpenAI's GPT-4.1 had already offered about 1 million tokens in 2025^[36]), surpassing the 400K tokens used by GPT-5.2 Thinking and matching what Anthropic and Google offered in their highest-tier 2026 models.^[10][12] The 128,000-token output cap is unchanged from the GPT-5 series, though it can be combined with the larger context window to support longer overall sessions through compaction.^[3] The August 31, 2025 knowledge cutoff is identical to GPT-5.2 and GPT-5.3, meaning GPT-5.4 has no awareness of post-cutoff events without retrieval or web search tools enabled.^[3]

What is native computer use?

Native computer use is the most novel addition. GPT-5.4 can take screenshots of a desktop, identify UI elements, move a virtual mouse, type into windows, and chain operations across applications.^[10] The OSWorld-Verified score of 75% indicates that its first-attempt completion rate on a curated set of 369 desktop tasks crossed the human reference baseline of 72.4%.^[1][12] In an evaluation by the customer Mainstay across roughly 30,000 HOA and property tax portals, GPT-5.4 achieved 95% first-attempt success and 100% within three attempts.^[1]

How does tool search cut token usage?

Tool search is the second major addition. Rather than including the full schema for every available tool in every API call, the model receives a lightweight catalog and queries it on demand.^[19][24] OpenAI reported a 47% reduction in token usage on benchmarks involving 36 MCP servers without loss of accuracy.^[19] The mechanism preserves the prompt cache better and scales to larger tool ecosystems than the prior "all tools in the prompt" approach.^[19]

How well does GPT-5.4 perform on benchmarks?

OpenAI published a wide set of benchmark scores at launch.^[1] Third-party trackers such as Artificial Analysis, Vellum, llm-stats.com, and DataCamp confirmed many of them, with the usual caveats that some scores reflect specific reasoning effort settings or tool configurations.^[8][10]

Headline benchmarks

The most prominent gains versus GPT-5.2 are on computer use and abstract reasoning. OSWorld-Verified rose from 47.3% to 75.0%, a 27.7 percentage-point jump.^[1][12] ARC-AGI-2 climbed from 52.9% to 73.3%, continuing the rapid progression that had begun with GPT-5.2. Knowledge work as measured by GDPval rose from 70.9% to 83.0%.^[1]

Coding benchmarks

GPT-5.4 absorbed the coding capability of GPT-5.3-Codex without surpassing it on every coding metric. SWE-bench Pro Public improved by 0.9 points (56.8% to 57.7%), and SWE-bench Verified stayed near the 80% range.^[14] Terminal-Bench 2.0 actually fell from 77.3% in GPT-5.3-Codex to 75.1% in GPT-5.4, reflecting the breadth-versus-specialization tradeoff of merging the Codex line back into the mainline.^[14] OpenAI's argument was that GPT-5.4's broader scope (long context, computer use, knowledge work, and tool search) made up for the marginal regression on terminal-specific tasks.^[1] Independent reviewers including Nathan Lambert (Interconnects) described the upgrade as "a meaningful step" in practice across correctness, ease of use, speed, and cost, even where on-paper benchmarks looked incremental.^[16]

Document and spreadsheet evaluations

On an internal OpenAI benchmark of spreadsheet modeling tasks of the kind a junior investment-banking analyst might perform, GPT-5.4 scored 87.3% mean compared to 68.4% for GPT-5.2.^[1] Human raters preferred GPT-5.4 presentations to GPT-5.2 presentations 68.0% of the time.^[1] On legal document analysis, legal AI company Harvey reported a 91% score on its BigLaw Bench evaluation.^[1]

How accurate is GPT-5.4?

OpenAI reported that individual factual claims in GPT-5.4 outputs are 33% less likely to be wrong than in GPT-5.2, and overall responses are 18% less likely to contain any factual errors.^[1] The measurement used de-identified prompts on which users had flagged factual errors.^[1]

Public leaderboards

On LM Arena, the news site VPSRanking reported that GPT-5.4 took the #1 Elo position in a March 6, 2026 update, the day after launch, though other leaderboard snapshots from March and April 2026 placed it behind Claude Opus 4.6 and Gemini 3.1 Pro.^[27] On Artificial Analysis's Intelligence Index it scored 57, tied with Gemini 3.1 Pro and ahead of Claude Opus 4.6 at 53.^[8] In Artificial Analysis measurements in April and May 2026, time to first token in reasoning mode ranged from about 170 to 230 seconds and throughput was about 80 to 86 tokens per second; the model used 120 million output tokens to complete the Intelligence Index evaluations.^[8]

How much does GPT-5.4 cost?

API pricing

The Thinking tier reflects a 43% increase in input price and a 7% increase in output price relative to GPT-5.2 Thinking.^[12] Cached input retains the 90% discount.^[12] Batch API pricing applies the standard 50% discount, and OpenAI's regional data residency endpoints add a 10% surcharge. The Decoder reported that GPT-5.4 mini and nano cost up to four times as much per token as GPT-5 mini and nano (mini is three times the input price of GPT-5 mini; nano is four times the input price of GPT-5 nano), a friction point for high-volume workloads.^[29]

GPT-5.4 Pro pricing puts it among the most expensive models on the market. Artificial Analysis consistently ranked it among the three most expensive models it tracked on both input and output price (147th of 149 on input and 148th of 149 on output in late May 2026).^[9]

ChatGPT plans

Free users initially had no access to GPT-5.4 itself, only the GPT-5.3 Instant model.^[1] The mini-tier addition on March 17 closed that gap by exposing GPT-5.4 mini through the Thinking feature on the free plan, although with rate limits.^[2][11]

What is GPT-5.4 used for?

Computer use

GPT-5.4's most-discussed feature is native computer use. OpenAI's announcement framed it as a step from chat-based assistance toward true desktop agency: the model can open applications, navigate menus, fill in spreadsheets, click through web forms, and chain operations without bespoke automation scripts.^[1][10] OpenAI's launch materials showed examples such as sending emails and scheduling calendar events through a browser interface, and customer Mainstay reported results on HOA and property tax portals.^[1]

Knowledge work

The GDPval score of 83% reflects performance across 44 occupations drawn from the top nine industries contributing to U.S. GDP.^[1] OpenAI's example workflows include producing a sales presentation, generating an accounting spreadsheet, scheduling staff for an urgent care clinic, drawing manufacturing diagrams, and producing short videos.^[1] The model is positioned for tasks that combine reading large documents, reasoning over them, drafting structured artifacts, and using tools to verify or refine the output.

Coding

With the merge of GPT-5.3-Codex's coding capability, GPT-5.4 became the recommended coding option in Codex CLI and the major IDE integrations.^[1][14] Reviewers noted improvements in long-running coding sessions: better context management, fewer "context wall" failures on million-token repositories, more reliable git operations, and reduced regression to previously solved problems.^[16] Cursor's evaluation data, cited by Nathan Lambert, showed efficiency gains in tokens per task.^[16] GPT-5.3-Codex retained narrow advantages in highly terminal-focused workflows where its specialist tuning still mattered.^[14]

Agentic workflows

The combination of long context, computer use, tool search, and the existing developer tools (apply_patch, local_shell, preambles) makes GPT-5.4 OpenAI's first model that can sustain agentic workflows over millions of tokens and many tools without specialist routing.^[1] Tool search reportedly cuts token usage on tool-heavy workflows by roughly half.^[19] Mid-response interactive thinking lets users redirect the model partway through a long reasoning trajectory rather than waiting for the entire run to finish.^[1]

Document, spreadsheet, and presentation generation

OpenAI emphasized professional artifacts in its launch demos: financial models, legal memos, board presentation decks, and multi-tab spreadsheets.^[1] The 87.3% spreadsheet score and the 68.0% presentation-preference rate are the headline numbers.^[1] Several customers quoted in OpenAI's launch post, including Mercor, Harvey, and Mainstay, cited improvements on their internal benchmarks.^[1]

How does GPT-5.4 compare with competitors?

At the time of GPT-5.4's release, the primary competing frontier models were Claude Opus 4.6 from Anthropic (Claude Opus 4.7 followed on April 16, 2026^[32]), Gemini 3.1 Pro from Google, and a small number of open-weights releases. The table below uses third-party benchmark numbers as reported in the months after launch.

GPT-5.4 led on computer use and knowledge work, sat near the top of the SWE-bench Verified pack, and was competitive on price for a frontier reasoning model.^{[8][12][31][33]} Claude Opus 4.6 retained narrow advantages on selected SWE-bench measurements and on tasks with ambiguous specifications, where reviewers found it inferred developer intent more reliably. Gemini 3.1 Pro held the top score on Humanity's Last Exam in the preview phase and offered cheaper input pricing along with better integration with Google Cloud.^[9]

The split between GPT-5.4 and competitors is qualitatively similar to the GPT-5.2 era: OpenAI's model is faster and more precise on well-specified tasks; Claude is more reliable on ambiguous ones; Gemini is strongest on long-context and Google-ecosystem workflows. GPT-5.4 closed the context-window gap that had persisted since the GPT-5 launch.

Is GPT-5.4 safe to use?

GPT-5.4 Thinking was the second model OpenAI classified as High capability for cybersecurity under the Preparedness Framework, after GPT-5.3-Codex earlier in 2026.^[22] OpenAI deployed it with an expanded version of the cyber safeguards introduced with GPT-5.3-Codex, including monitoring systems, trusted access controls, and asynchronous blocking of higher-risk requests for customers on Zero Data Retention surfaces.^[1]

The High threshold under the framework is defined as a model that "removes existing bottlenecks to scaling cyber operations" through automation of end-to-end attacks against hardened targets or automation of operationally relevant vulnerability discovery and exploitation.^[28] OpenAI said it was treating GPT-5.4 Thinking as High "even though we cannot be certain that it actually has these capabilities," because the model met the requirements of each of its canary thresholds.^[28]

OpenAI also reported safety improvements for ordinary user interactions. In the system card, GPT-5.4 Thinking outperformed previous models on all of OpenAI's dynamic mental-health evaluations.^[28]

GPT-5.4-Cyber, released on April 14, 2026, formalized the cyber-permissive variant for verified defensive professionals.^[21][22] OpenAI deprecated it on September 11, 2026 and set its API removal for October 1, 2026.^[30] It includes binary reverse engineering capability, lower refusal rates on legitimate vulnerability research, and a tiered access model with identity verification, enterprise authentication, and Zero-Data-Retention waivers for the highest permissiveness tier.^[21][22]

How was GPT-5.4 received?

Developer and enterprise reception

Developer reception was largely positive on capability. Reviewers from Turing College, NxCode, BuildFastWithAI, and Interconnects characterized GPT-5.4 as a meaningful step forward, particularly on long sessions where context-window pressure had been a persistent bottleneck for the GPT-5.2 generation.^{[13][16][17][18]} Cursor's data showed efficiency gains in tokens per coding task.^[16] Mercor named GPT-5.4 the leader on its APEX-Agents law and finance benchmark.

Nathan Lambert summarized the trajectory: "Where GPT 5.4 feels like another incremental model on some on-paper benchmarks, in practice it feels like a meaningful step" across correctness, ease of use, speed, and cost.^[16] He praised the elimination of the "death by a thousand cuts" experience around git operations and background package management that had plagued earlier Codex variants.^[16]

Enterprise customers cited workflow improvements. Mainstay reported 95% first-attempt success on HOA and property tax portals, and Mercor, Harvey, and Notion were among the customers quoted in OpenAI's launch post.^[1] Microsoft's day-zero Azure AI Foundry availability gave enterprise users a fast path to deployment.

Criticism

Criticism focused on three areas. First, pricing: the 43% input-price increase over GPT-5.2 Thinking, and price increases of up to four times for mini and nano relative to the GPT-5 small models, drew complaints from cost-sensitive teams.^[12][29] The Decoder ran a critical piece on the small-model price increases.^[29]

Second, latency: Artificial Analysis measured times to first token of roughly 170 to 230 seconds in reasoning mode during April and May 2026, which is partly inherent to extended reasoning.^[8]

Third, mixed retention of Codex specialization: the Terminal-Bench 2.0 regression from 77.3% to 75.1% relative to GPT-5.3-Codex prompted some terminal-heavy teams to keep GPT-5.3-Codex as their primary coding model.^[14] GPT-5.4's broader scope was seen as a tradeoff against the highly tuned specialist model.^[14]

Reaction in coverage and the open community

Coverage in outlets including TechCrunch and WIRED framed GPT-5.4 around professional work and the AI coding-agent race, with the ChatGPT for Excel add-in and the promise of fewer iteration loops on professional artifacts drawing attention.^[6][20] Reactions on Reddit and Hacker News were mixed, with some users describing an "oversmart" feel that made the model harder to steer.^[20]

What came after GPT-5.4?

GPT-5.4 was succeeded by GPT-5.5, released April 23, 2026, seven weeks after the GPT-5.4 launch. GPT-5.5 carried forward the 1M-token context window and computer-use capability, increased token efficiency, raised intelligence-index scores further, and added new agentic features.^[21] GPT-5.5 was priced higher than GPT-5.4 in the API but with lower effective per-task cost on many workloads due to fewer tokens consumed. GPT-5.5 Instant followed on May 5, 2026 as the default model in ChatGPT.

As of November 2026, OpenAI's deprecations page listed no shutdown date for gpt-5.4 itself, while gpt-5.4-nano had been deprecated on October 1, 2026 (with API removal set for April 1, 2027) and gpt-5.4-cyber had been removed from the API on October 1, 2026.^[30] OpenAI had since released further models, including GPT-5.6 and GPT-6 Astra.

Continue reading

• GPT-5GPT-5 is a family of proprietary large language models released by OpenAI on August 7, 2025.Mentioned 169 times above
• OpenAIOpenAI is an American artificial intelligence research and deployment organization.Mentioned 45 times above
• GPT-5.3GPT-5.3 is a family of large language models released by OpenAI in February and March 2026 as part of the GPT-5 lineage.Mentioned 20 times above
• GPT-5.2GPT-5.2 is a large language model developed by OpenAI and released on December 11, 2025. It is the second major update to the GPT-5 series, succeeding GPT-5.1, which had been released about a month earlier.Mentioned 16 times above
• ChatGPTChatGPT is a conversational artificial intelligence chatbot developed by OpenAI and built on the company's GPT series of large language models.Mentioned 12 times above
• SSWE-benchSWE-bench is an execution-based benchmark for evaluating whether a language-model system can resolve real software issues.Mentioned 7 times above

More to read

• Gemini 3
• GPT-5.5
• GPT-5.1
• GPT-4
• GPT-4.1
• Large Language Model

Topics:AI ModelsLarge Language ModelsOpenAI

See also

• GPT-5
• GPT-5.1
• GPT-5.2
• GPT-5.3
• GPT-5.5
• GPT-4
• GPT-4.1
• GPT-4o
• o1
• o3
• ChatGPT
• OpenAI
• Sam Altman
• Large language model
• SWE-bench
• GPQA
• AIME
• MMLU
• MMMU
• HumanEval
• Tau-bench
• LM Arena
• Claude Opus 4.7 (1M context)
• Gemini 3
• AI API cost and context planner
• AI model comparison finder
• AI model lifecycle calendar

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40 ^41 ^42 ^43 ^44OpenAI. "Introducing GPT-5.4." openai.com, March 5, 2026. openai.com/...introducing-gpt-5-4
2. ^1 ^2 ^3 ^4 ^5 ^6OpenAI. "Introducing GPT-5.4 mini and nano." openai.com, March 17, 2026. openai.com/...introducing-gpt-5-4-mini-and-nano
3. ^1 ^2 ^3 ^4 ^5OpenAI. "GPT-5.4 Model." developers.openai.com. developers.openai.com/...gpt-5.4
4. ^OpenAI. "GPT-5.4 mini Model." developers.openai.com. developers.openai.com/...gpt-5.4-mini
5. ^1 ^2OpenAI. "GPT-5.4 nano Model." developers.openai.com. developers.openai.com/...gpt-5.4-nano
6. ^1 ^2 ^3 ^4TechCrunch. "OpenAI launches GPT-5.4 with Pro and Thinking versions." techcrunch.com, March 5, 2026. techcrunch.com/...4-with-pro-and-thinking-versions
7. Wikipedia contributors. "GPT-5.4." Wikipedia. en.wikipedia.org/...GPT-5.4
8. ^1 ^2 ^3 ^4 ^5 ^6Artificial Analysis. "GPT-5.4 (xhigh) Intelligence, Performance & Price Analysis." artificialanalysis.ai. artificialanalysis.ai/...gpt-5-4
9. ^1 ^2 ^3 ^4Artificial Analysis. "GPT-5.4 Pro (xhigh) Intelligence, Performance & Price Analysis." artificialanalysis.ai. artificialanalysis.ai/...gpt-5-4-pro
10. ^1 ^2 ^3 ^4 ^5 ^6DataCamp. "GPT-5.4: Native Computer Use, 1M Context Window, Tool Search." datacamp.com. datacamp.com/...gpt-5-4
11. ^1 ^2 ^3DataCamp. "GPT-5.4 mini and nano: Benchmarks, Access, and Reactions." datacamp.com. datacamp.com/...gpt-5-4-mini-nano
12. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8NxCode. "GPT-5.4 (March 2026): 75% Computer Use, 1M Context, $2.50/MTok." nxcode.io. nxcode.io/...-4-release-date-features-pricing-2026
13. ^NxCode. "GPT 5.4 Complete Guide 2026: Features, Pricing, Benchmarks & How to Use." nxcode.io. nxcode.io/...te-guide-features-pricing-models-2026
14. ^1 ^2 ^3 ^4 ^5 ^6 ^7NxCode. "GPT-5.4 vs GPT-5.3 Codex: Should Developers Upgrade?." nxcode.io. nxcode.io/...gpt-5-3-codex-upgrade-comparison-2026
15. ^1 ^2NxCode. "OpenAI GPT-5 Model Guide: GPT-5.2 vs 5.3 vs 5.4." nxcode.io. nxcode.io/...i-gpt-5-model-guide-which-to-use-2026
16. ^1 ^2 ^3 ^4 ^5 ^6 ^7Lambert, Nathan. "GPT 5.4 is a big step for Codex." interconnects.ai. interconnects.ai/...gpt-54-is-a-big-step-for-codex
17. ^BuildFastWithAI. "GPT-5.4 Review: Features, Benchmarks & Access (2026)." buildfastwithai.com. buildfastwithai.com/...-5-4-review-benchmarks-2026
18. ^Turing College. "GPT-5.4 Review: Is It Worth Leaving GPT-5.3 Codex Behind?." turingcollege.com. turingcollege.com/...gpt-5-4-review-vs-gpt-5-3-codex
19. ^1 ^2 ^3 ^4MindStudio. "What Is Tool Search? How GPT-5.4 Cuts Token Usage by 47%." mindstudio.ai. mindstudio.ai/...l-search-gpt-5-4-token-efficiency
20. ^1 ^2AI Critique. "GPT-5.4 and the March 2026 ChatGPT Upgrade Cycle." aicritique.org, March 16, 2026. aicritique.org/...ratives-and-real-world-reactions
21. ^1 ^2 ^3 ^4 ^5 ^6 ^7Fluid Attacks. "GPT-5.4-Cyber and GPT-5.5 for security." fluidattacks.com. fluidattacks.com/...pt-5-5-ai-cybersecurity-future
22. ^1 ^2 ^3 ^4 ^5 ^6OpenAI. "Trusted access for the next era of cyber defense." openai.com. openai.com/...ing-trusted-access-for-cyber-defense
23. ALMCorp. "OpenAI GPT-5.4: Features, Benchmarks, Pricing & Computer Use (2026)." almcorp.com. almcorp.com/...gpt-5-4
24. ^1 ^2CometAPI. "How to Use GPT-5.4 API: Parameters and Tools Usage Guide." cometapi.com. cometapi.com/how-to-use-gpt-5-4-api
25. TTMS. "GPT-5.4 by OpenAI: What's new? 9 Key Improvements." ttms.com. ttms.com/...by-openai-whats-new-9-key-improvements
26. AnalyticsVidhya. "We Tried The New GPT-5.4 And it is The Most Powerful ChatGPT Has Ever Been." analyticsvidhya.com, March 2026. analyticsvidhya.com/...we-tried-gpt-5-4-ai
27. ^1 ^2VPSRanking. "LMArena Ranking Update GPT-5.4 Takes #1 Elo Rating." vpsranking.com. vpsranking.com/...ai-2026-03-06-lmarena
28. ^1 ^2 ^3 ^4OpenAI. "GPT-5.4 Thinking System Card." deploymentsafety.openai.com, March 5, 2026. deploymentsafety.openai.com/gpt-5-4-thinking
29. ^1 ^2 ^3Bastian, Matthias. "OpenAI ships GPT-5.4 mini and nano, faster and more capable but up to 4x pricier." The Decoder, March 17, 2026. the-decoder.com/...re-capable-but-up-to-4x-pricier
30. ^1 ^2OpenAI. "Deprecations." developers.openai.com. developers.openai.com/...deprecations
31. ^1 ^2Anthropic. "Introducing Claude Opus 4.6." anthropic.com, February 5, 2026. anthropic.com/...claude-opus-4-6
32. ^Anthropic. "Introducing Claude Opus 4.7." anthropic.com, April 16, 2026. anthropic.com/...claude-opus-4-7
33. ^Google DeepMind. "Gemini 3.1 Pro - Model Card." deepmind.google. deepmind.google/...gemini-3-1-pro
34. ^GitHub. "GPT-5.4 is generally available in GitHub Copilot." github.blog, March 5, 2026. github.blog/...nerally-available-in-github-copilot
35. ^Microsoft. "Introducing GPT-5.4 in Microsoft Foundry." techcommunity.microsoft.com, March 5, 2026. techcommunity.microsoft.com/...4499785
36. ^OpenAI. "GPT-4.1 Model." developers.openai.com. developers.openai.com/...gpt-4.1

Show all 36 references