GPT-5, often searched as ChatGPT 5, is the OpenAI model released on August 7, 2025 (API snapshot gpt-5-2025-08-07). As of October 2026 it is no longer OpenAI's current model. OpenAI's model page calls GPT-5 its previous model and recommends GPT-6 Astra, and OpenAI's deprecations page schedules the GPT-5 API snapshots for removal on December 11, 2026. This guide covers GPT-5's release date, launch features and benchmarks, what replaced it, and how teams use current GPT models in Slack.
TL;DR. GPT-5 launched August 7, 2025 with a 400,000-token context window and API pricing of $1.25 per million input tokens and $10 per million output tokens. In September 2026 OpenAI released the GPT-6 family (GPT-6 Astra, GPT-6 Sol, GPT-6 Luna, then GPT-6.1 Sol), and the gpt-5, gpt-5-mini, gpt-5-nano, and gpt-5-pro snapshots leave the API on December 11, 2026. Facts as of October 2026.
GPT-5 release date and status (as of October 2026)
- Released: August 7, 2025. The API launched with gpt-5, gpt-5-mini, and gpt-5-nano; a gpt-5-pro snapshot followed on October 6, 2025 (OpenAI deprecations).
- Launch specs (API): 400,000-token context window, 128,000 max output tokens, knowledge cutoff September 30, 2024 (OpenAI GPT-5 model page).
- Launch pricing (gpt-5): $1.25 per million input tokens, $0.125 cached input, $10 per million output tokens (OpenAI GPT-5 model page).
- Retirement: on June 11, 2026 OpenAI notified developers that the GPT-5 snapshots leave the API on December 11, 2026. Recommended replacements are gpt-5.6-sol for gpt-5 and gpt-5-pro, gpt-5.6-terra for gpt-5-mini, and gpt-5.6-luna for gpt-5-nano. The gpt-5-chat-latest alias was already shut down on July 23, 2026 (OpenAI deprecations).
What replaced GPT-5: the GPT-6 family
OpenAI's models page now lists three flagship models. Release dates and prices below are from the OpenAI API changelog, as of October 2026:
- GPT-6 Astra (gpt-6-astra), released September 3, 2026: OpenAI's most capable model, aimed at complex reasoning and coding.
- GPT-6 Sol and GPT-6 Luna, released September 22, 2026: Sol at $2 input and $10 output per million tokens, Luna at $0.10 input and $0.50 output.
- GPT-6.1 Sol (gpt-6.1-sol), released September 29, 2026: near-Astra performance for complex work at a lower cost than Astra.
- GPT-6.1 Terra (gpt-6.1-terra), released October 26, 2026: $1 input and $5 output per million tokens, half of Sol's rate. For a side-by-side of the two GPT-6.1 models, see GPT-6.1 Sol vs GPT-6.1 Terra.
Inside the ChatGPT app, the model each plan gets changes more often than the API lineup. OpenAI's help center article "GPT-6 and other models in ChatGPT" is the place to check current plan defaults.
The rest of this post documents GPT-5 as it launched in August 2025, which is still useful if you are comparing models or migrating off a GPT-5 snapshot.
What Is GPT-5?
At launch, GPT-5 was OpenAI's most advanced language model, built to be more robust, reliable, and helpful than its predecessors. OpenAI reported state-of-the-art accuracy on real-world tasks along with fewer hallucinations and deceptive responses.
Key Improvements in GPT-5
- State-of-the-art performance on leading AI benchmarks:
- Fewer hallucinations: GPT-5 made fewer factual errors than GPT-4o and OpenAI o3 in OpenAI's launch tests.
- Safer and more honest: Reduced deceptive behavior in edge cases or underspecified prompts.
- Improved instruction following: More reliable for agents, workflows, and multi-step tasks.
- Multimodal reasoning: Enhanced capabilities across text, image, video, and charts.
What makes GPT-5 different?
Rather than a single monolithic model, GPT-5 introduces a router-based model architecture:
- GPT-5: Fast, general-purpose response generation
- GPT-5 Thinking: Deep reasoning for complex tasks
- GPT-5 Pro: Extended reasoning for the hardest tasks
These models are deployed intelligently behind the scenes, depending on query difficulty, user intent, and context. This allows GPT-5 to feel both fast and deeply capable, delivering expert-level answers where needed, while still handling everyday queries with speed and grace.
Why GPT-5 Feels More Trustworthy
By OpenAI's measurements, GPT-5 also behaved more responsibly than earlier models: it explained its limits better and made fewer false claims.
Fewer Hallucinations, Even When Reasoning
GPT-5 is significantly more factually accurate:
- About 45% fewer factual errors than GPT-4o, and about 80% fewer than OpenAI o3 with thinking enabled
- Over 5x fewer factual errors than OpenAI o3 on open-ended benchmarks like LongFact and FActScore
Whether answering complex science questions or writing long-form content, GPT-5 produces more grounded, verifiable answers.
More Honest When Tasks Can't Be Done
GPT-5 is trained to say “I don't know” when the task is impossible, underspecified, or missing key inputs:
- In OpenAI's deception evaluations, a monitor flagged deceptive responses in about 2.1% of GPT-5 thinking responses, down from 4.8% for o3
- In multimodal tests with missing images, GPT-5 hallucinated only 9% of the time, compared to o3's 86.7%
This means better transparency and fewer misleading answers.
Safety with Nuance
With GPT-5, OpenAI introduced safe completions, a new training paradigm that:
- Avoids over-refusing ambiguous prompts
- Answers safely when possible
- Explains why a task can't be done, instead of deflecting
This enables GPT-5 to be safer without feeling frustratingly limited.
Less Sycophantic, More Thoughtful
GPT-5 feels less like a cheerleader and more like a thoughtful colleague:
- Sycophantic replies reduced by over 50%
- Clearer follow-ups and fewer unnecessary flattery
Conversations become more productive and grounded.
Evaluations: How Smart Is GPT-5?
Math Mastery
GPT-5 dominates math evaluations:
- AIME 2025: 94.6% (no tools), a new state of the art.
- HMMT: 96.7% (no tools), 100% (with tools).
- FrontierMath: 26.3% → 32.1% with tool support.
- GPQA (PhD-level): 88.4% (no tools) → 89.4% (with thinking).
Real-World Coding
On practical software engineering and code editing benchmarks:
- SWE-bench Verified: 74.9% accuracy (GPT-5) vs 69.1% (OpenAI o3) and 30.8% (GPT-4o).
- Aider Polyglot: 88% accuracy (GPT-5), far ahead of all previous models.
These improvements suggest GPT-5 can handle significantly more complex engineering workflows.
Instruction Following & Tool Use
GPT-5 is vastly better at multi-step reasoning and agentic behaviors:
- Scale MultiChallenge (multi-turn): 69.6% vs GPT-4o's 40.3%.
- BrowseComp (search + browsing): 54.9% vs 49.7%.
- COLLIE (freeform following): 99.0% accuracy.
It not only follows instructions better but coordinates tools to complete tasks more reliably.
Multimodal Understanding
GPT-5 outperforms on visual, video, and diagram-based tasks:
- MMMU: 84.2% vs GPT-4o's 72.2%.
- ERQA: 65.7% vs 35.2% (GPT-4o).
- CharXiv Reasoning: 81.1% vs 58.8%.
This means stronger capabilities for interpreting images, visual documents, and presentations.
Health Conversations
GPT-5 is the most accurate and least hallucinatory model for medical applications:
- HealthBench: 67.2% vs GPT-4o's 32.0%.
- Hallucination Rate: 1.6% (with thinking), down from GPT-4o's 15.8%.
Its performance on sensitive conversations makes it viable for high-trust settings.
Economically Valuable Tasks
GPT-5 beats both o3 and ChatGPT Agent on complex, real-world professional tasks:
- Outperforms in law, logistics, sales, and engineering.
- Achieves 47.1% wins over industry experts in internal benchmarks.
This suggests GPT-5 is not just smarter, it's more useful in economically important domains.
Faster, More Efficient Thinking
GPT-5 achieves better performance with fewer output tokens than OpenAI o3:
- 50–80% fewer tokens across reasoning, coding, and scientific benchmarks.
- Better performance with less effort and latency, critical for real-time use.
Detailed Benchmarks
Who is GPT-5 for?
- Business leaders & teams: Analyze documents, monitor operations, and plan strategy with better context awareness and safer outputs.
- Developers: Use GPT-5 to generate production-quality frontends, debug codebases, and automate software tasks.
- Knowledge workers & writers: Get help crafting thoughtful reports, articles, or even poetry, with deeper structure and emotional impact.
Using GPT models with your team in Slack
Most teams asking about GPT in Slack do not want each person pasting questions into a personal chat. They want one shared agent the channel can @mention, grounded in the team's own knowledge and tools.
In Runbear the model is a per-agent setting. Claude Agent SDK agents can use Claude, GPT, or Gemini models, and OpenAI Responses agents use GPT models. On Claude Agent SDK agents the GPT rows are named by tier (Astra, Sol, Terra, Luna) and carry a Stable badge that follows the latest available version in that tier; today that is GPT-6 for Astra and Luna, GPT-6.1 for Sol, and GPT-5.6 for Terra. To keep an agent on one exact release, choose it under Specific version. See the agent settings docs.
Runbear agents work in Slack and Microsoft Teams. In Slack they always respond to @mentions and direct messages; replying without an @mention or on keywords is an extra switch you turn on per channel.
If you still call a gpt-5 snapshot directly through the API, plan the move to a supported model before December 11, 2026.
Related guides: how to add an AI agent to a Slack channel, Runbear vs ChatGPT for Slack, Claude Haiku 5.5 in Slack, and Gemini 4 Argon in Slack. If you are weighing OpenAI's personal agent, see ChatGPT Dots alternatives for Slack and Microsoft Teams. On Microsoft Teams instead? See Microsoft Teams ChatGPT integration: @ChatGPT, dots, or a shared agent.
Final Thoughts
GPT-5 raised the bar on reasoning, coding, and hallucination rates when it launched in August 2025. Fourteen months later it is a previous-generation model: the GPT-6 family is OpenAI's current lineup, and the GPT-5 API snapshots shut down on December 11, 2026. For teams, the practical question is less which model leads this month and more whether your agent can move to the next one without being rebuilt.
Frequently Asked Questions
When was GPT-5 released?
OpenAI released GPT-5 on August 7, 2025, matching the API snapshot name gpt-5-2025-08-07. The gpt-5-pro-2025-10-06 snapshot followed later (OpenAI deprecations).
Is GPT-5 still available in 2026?
In the API, the gpt-5, gpt-5-mini, gpt-5-nano, and gpt-5-pro snapshots remain until December 11, 2026, when OpenAI removes them. The gpt-5-chat-latest alias was shut down on July 23, 2026. OpenAI's GPT-5 model page recommends GPT-6 Astra, and the deprecations page lists a GPT-5.6 replacement for each snapshot.
What replaced GPT-5?
For the API, OpenAI's recommended replacements are gpt-5.6-sol (for gpt-5 and gpt-5-pro), gpt-5.6-terra (for gpt-5-mini), and gpt-5.6-luna (for gpt-5-nano). OpenAI's newest flagship models are GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna, released in September 2026 (OpenAI API changelog).
How much does GPT-5 cost?
At launch, gpt-5 cost $1.25 per million input tokens, $0.125 per million cached input tokens, and $10 per million output tokens (OpenAI GPT-5 model page). For comparison, GPT-6 Sol is $2 input and $10 output per million tokens (OpenAI API changelog). ChatGPT plan prices are listed on OpenAI's pricing page.
How is GPT-5 different from GPT-4o?
By OpenAI's launch figures, GPT-5 made about 45% fewer factual errors than GPT-4o, and with thinking about 80% fewer than OpenAI o3. It scored 69.6% on the Scale MultiChallenge multi-turn benchmark versus GPT-4o's 40.3%, and used 50 to 80% fewer output tokens than OpenAI o3 for better results. In ChatGPT it also introduced a router that picks a deeper thinking mode for harder questions.
How does GPT-5 handle hallucinations?
OpenAI reported that GPT-5 with thinking was about 80% less likely than OpenAI o3 to make a factual error, and that a monitor flagged deceptive responses in about 2.1% of GPT-5 thinking responses versus 4.8% for o3. On HealthBench its hallucination rate was 1.6% with thinking, versus 15.8% for GPT-4o.
Can I use GPT models in Slack?
Yes. With Runbear, a Claude Agent SDK agent can use GPT models chosen by tier (Astra, Sol, Terra, Luna), and the Stable option follows the latest version in that tier. The agent answers @mentions and DMs in Slack and also works in Microsoft Teams (agent settings docs).
Get StartedLaunch benchmark scores and images in this post are from OpenAI's GPT-5 announcements: Introducing GPT-5 and Introducing GPT-5 for developers. Specs, pricing, and retirement dates are from the OpenAI GPT-5 model page, OpenAI deprecations, and OpenAI API changelog, as of October 2026.
