diff --git a/.optimize-cache.json b/.optimize-cache.json index 8df30d88bd..2e03da9dc6 100644 --- a/.optimize-cache.json +++ b/.optimize-cache.json @@ -699,7 +699,7 @@ "static/images/blog/google-oauth-expo/login-successful.png": "441ef69e3ba8716c4ff4c20dc8c31f8f3f418010a11ff1e74585645bad414f74", "static/images/blog/gpt-5-5-launch/card-problem.png": "9cfd1fa44c5f8d76a54904d1e656595f1580721d498952a96105dd6f9ddc82fd", "static/images/blog/gpt-5-5-launch/cover.png": "768138556cc8a3c1645ff1c602ea1ac5ba2eb373f8f642d2c023e6380950ec08", - "static/images/blog/gpt-56-is-here-openais-efficient-frontier-model/cover.png": "2156c5047fa2630e148ea0fd5189c4d08579b5c96e92530c4ae45b642acb6589", + "static/images/blog/gpt-56-is-here-openais-efficient-frontier-model/cover.png": "113febf4d6a1f3403b7a633411678525234c8366184d5f224a98265353d525ab", "static/images/blog/grok-45-coding-model/cover.png": "b40f87952b940767f393a93a7628f8277119862a3a3ac7b7d40d7ee5679213ff", "static/images/blog/hacktoberfest-2023.png": "e0521ba26e59dacf91072391f7eb0df0e724457961fec4cb04a4914b31dd0024", "static/images/blog/hacktoberfest2023-swag.png": "7a0af270f7fe24c53e2497ca59bead84255d2d2e882dd38473b1bc6bb0a22d1f", diff --git a/src/routes/blog/post/gpt-56-is-here-openais-efficient-frontier-model/+page.markdoc b/src/routes/blog/post/gpt-56-is-here-openais-efficient-frontier-model/+page.markdoc index 1036cb5e85..a3020ed378 100644 --- a/src/routes/blog/post/gpt-56-is-here-openais-efficient-frontier-model/+page.markdoc +++ b/src/routes/blog/post/gpt-56-is-here-openais-efficient-frontier-model/+page.markdoc @@ -1,171 +1,123 @@ --- layout: post -title: "GPT-5.6 is here: OpenAI's efficient frontier model" -description: "GPT-5.6 is OpenAI's new frontier model family with Sol, Terra, and Luna. See how it performs on coding and agents, plus what it means for developers." -date: 2026-07-10 -cover: /images/blog/gpt-56-is-here-openais-efficient-frontier-model/cover.avif -timeToRead: 5 +title: "GPT-5.6 is here: OpenAI's efficient coding models" +description: "GPT-5.6 brings OpenAI's Sol, Terra, and Luna coding models to the API. Explore their benchmarks, pricing, agent tools, and safety trade-offs." +date: 2026-07-15 +cover: /images/blog/gpt-56-is-here-openais-efficient-frontier-model/cover.png +timeToRead: 7 author: aishwari category: ai featured: false +unlisted: true faqs: - - question: What is the difference between Sol, Terra, and Luna? - answer: Sol is the flagship for the highest capability, Terra is a lower-cost model with performance competitive with GPT-5.5, and Luna is the fastest and most affordable. They are durable capability tiers, so each can advance on its own cadence. - - question: How much does GPT-5.6 cost? - answer: Per 1M tokens, Sol is $5 input and $30 output, Terra is $2.50 and $15, and Luna is $1 and $6. Prompt caching adds cache writes at 1.25x the uncached input rate and cache reads at a 90% discount. - - question: Is GPT-5.6 good for coding? - answer: Yes. GPT-5.6 Sol sets a new state of the art on the Artificial Analysis Coding Agent Index at 80 and leads on Terminal-Bench 2.1 and DeepSWE, though Claude Mythos 5 still leads SWE-Bench Pro. Its main edge is doing this with far fewer tokens and less time. - - question: Where can I use GPT-5.6? - answer: It is available across ChatGPT, Codex, and the OpenAI API, rolling out globally over 24 hours from launch, with access varying by plan and surface. + - question: "What are GPT-5.6 Sol, Terra, and Luna?" + answer: "They are three capability tiers in OpenAI's GPT-5.6 family. Sol is the flagship, Terra balances capability and cost, and Luna is the fastest and least expensive tier." + - question: "How much does the GPT-5.6 API cost?" + answer: "Per 1 million tokens, Sol costs $5 for input and $30 for output, Terra costs $2.50 and $15, and Luna costs $1 and $6. Cache writes cost 1.25 times the uncached input rate, while cache reads receive a 90% discount." + - question: "Is GPT-5.6 better than Claude Fable 5 for coding?" + answer: "It depends on the evaluation. GPT-5.6 Sol leads the independent Artificial Analysis Coding Agent Index with 80, ahead of Claude Fable 5 at 77.2, but Claude Fable 5 leads OpenAI's reported SWE-Bench Pro comparison, 80% to 64.6%." + - question: "What is GPT-5.6 ultra mode?" + answer: "Ultra is OpenAI's highest-capability setting. It coordinates parallel agents on separate workstreams, then combines their results. It uses more compute than single-agent settings and is intended for demanding tasks where latency and quality justify the added cost." + - question: "Can developers use GPT-5.6 through the API?" + answer: "Yes. Sol, Terra, and Luna are available through the OpenAI API. The Responses API also supports Programmatic Tool Calling, while concurrent multi-agent execution launched in beta." --- -OpenAI released the [GPT-5.6 family](https://openai.com/index/gpt-5-6/) for general availability on Thursday, July 9, 2026, following a limited preview. The launch is easy to summarize in one line: the same or better quality than the previous generation, at lower cost and faster time to result. This post covers what GPT-5.6 is, what is genuinely new, how it performs, how it is already changing the way OpenAI works internally, and what it means if you build software. -# What is GPT-5.6? +Coding benchmarks are becoming harder to read at a glance. A higher score can come from a more capable model, a better agent harness, more reasoning time, parallel workers, or simply more tokens. OpenAI's new [GPT-5.6 family](https://openai.com/index/gpt-5-6/) makes that distinction especially important because its main claim is not raw capability alone. It is **more completed work per token and per dollar**. -GPT-5.6 is OpenAI's latest generation of frontier models, launched on July 9, 2026. It ships as three tiers designed to be efficient by default, spending fewer tokens and less time to reach comparable or better results than GPT-5.5, with higher effort settings available when a task needs them. +The release, highlighted in the [source post on X](https://x.com/bruteforcearete/status/2075592935664660876), introduces three API models named **Sol, Terra, and Luna**. It also brings Programmatic Tool Calling and beta multi-agent execution to the Responses API, plus an `ultra` setting that coordinates parallel agents for difficult work. -The number identifies the generation. Sol, Terra, and Luna are durable capability tiers that can advance on their own cadence. +# GPT-5.6 Sol, Terra, and Luna target different workloads -| Model | Position | Price per 1M tokens (input / output) | -| ------------- | ------------------------------------ | ------------------------------------ | -| GPT-5.6 Sol | Flagship, highest capability | $5 / $30 | -| GPT-5.6 Terra | Lower cost, competitive with GPT-5.5 | $2.50 / $15 | -| GPT-5.6 Luna | Fastest and most affordable | $1 / $6 | +OpenAI describes the names as durable capability tiers rather than one-off model labels. Sol is the flagship, Terra is the balanced option, and Luna prioritizes speed and cost. That gives developers a stable way to choose a tier even as each one advances on its own release cadence. -# What is actually new in GPT-5.6? +| Model | Intended role | API input per 1M tokens | API output per 1M tokens | +| --- | --- | ---: | ---: | +| GPT-5.6 Sol | Highest capability | $5 | $30 | +| GPT-5.6 Terra | Balanced capability and cost | $2.50 | $15 | +| GPT-5.6 Luna | Fastest and most affordable | $1 | $6 | -The headline is efficiency. Instead of forcing you to pick a slow, expensive model for hard problems and a cheap one for easy problems, GPT-5.6 is tuned to do more with each token by default, then scale up only when you ask it to. +Those list prices do not describe the full cost of an agent run. A model that finishes in fewer turns, makes fewer tool calls, or emits fewer reasoning tokens can cost less even with a higher per-token rate. Conversely, `max` and `ultra` can spend more compute to improve the chance of finishing a hard task. -The clearest evidence shows up across two independent evaluations: +GPT-5.6 also changes prompt caching economics. OpenAI charges cache writes at **1.25 times the normal input rate**, retains a **90% discount for cache reads**, supports explicit cache breakpoints, and guarantees a minimum cache life of 30 minutes. Applications with a large stable system prompt or tool catalog should model both the initial write and subsequent reuse instead of treating all cached input as equally discounted. -* **Agents' Last Exam:** OpenAI highlights a score of 53.6 for GPT-5.6 Sol in its launch announcement, while its full benchmark table lists 52.7%. Both results place it ahead of Claude Fable 5, although the difference appears to reflect different evaluation configurations. -* **Efficiency across the family:** At medium reasoning, Sol beats Fable 5 by 11.4 points at roughly one quarter of the estimated cost. Terra and Luna outperform Fable 5 at around one sixteenth of the cost. -* **Artificial Analysis Intelligence Index:** Sol with max reasoning comes within one point of Fable 5 while completing tasks in 61% less time at roughly half the cost. +# GPT-5.6 coding benchmarks show strength and trade-offs -Two mechanisms make that flexibility possible. +The clearest coding result comes from [Artificial Analysis's independent evaluation](https://artificialanalysis.ai/articles/gpt-5-6-has-landed/). Its Coding Agent Index combines DeepSWE, Terminal-Bench, and SWE-Atlas-QnA, with each model paired with an agent harness. GPT-5.6 Sol at `max` reasoning, running in Codex, leads the index with a score of 80. -**Effort settings that trade cost for capability.** +| Model and configuration | Coding Agent Index | DeepSWE v1.1 | Terminal-Bench 2.1 | +| --- | ---: | ---: | ---: | +| GPT-5.6 Sol (`max`) | 80 | 72.7% | 88.8% | +| GPT-5.6 Terra (`max`) | 77.4 | 69.6% | 87.4% | +| GPT-5.6 Luna (`max`) | 74.6 | 67.2% | 84.7% | +| GPT-5.5 | 76.4 | 67% | 85.6% | +| Claude Fable 5 | 77.2 | 69.7% | 83.1% | +| Claude Opus 4.8 | 72.5 | 59% | 78.9% | -* `max` gives the model more time than the previous `xhigh` setting to reason, run checks, and revise its approach. -* `ultra` goes further, coordinating four agents in parallel by default, trading higher token use for stronger results and faster time to result on demanding tasks. Across BrowseComp, SEC-Bench Pro, and Terminal-Bench 2.1, adding parallel agents pushes the score-to-latency frontier up and to the left, reaching stronger results in less time. +The benchmark values above come from [OpenAI's published comparison table](https://openai.com/index/gpt-5-6/) and Artificial Analysis. Artificial Analysis reports rounded index scores of 77 and 75 for Terra and Luna, while OpenAI publishes 77.4 and 74.6. Sol also led each component in Artificial Analysis's run, tying Grok 4.5 on SWE-Atlas-QnA. -**Tooling built for agents.** +Efficiency is the more useful result. Artificial Analysis measured Sol at about **40% lower cost per coding task than Claude Fable 5** and about 10% lower than Claude Opus 4.8 in their respective harnesses. Terra and Luna reduced per-task cost by roughly 60% and 80% compared with Sol. On the broader Intelligence Index, Sol used about 15,000 output tokens per task, compared with 16,000 for GPT-5.5. -* [**Programmatic Tool Calling**](https://platform.openai.com/docs/guides/tools) in the Responses API lets GPT-5.6 write and run lightweight programs that coordinate tools, filter large amounts of intermediate data, keep only what matters, and adapt as work unfolds. Instead of scripting every step or passing every tool response back through the model, tool-heavy tasks advance with fewer tokens, fewer model round trips, and less hand-holding. -* **Multi-agent**, in beta in the Responses API, lets GPT-5.6 run concurrent subagents and synthesize their work in a single request. This is the same mechanism behind `ultra`. +No single table establishes a universal coding winner. In OpenAI's SWE-Bench Pro results, Claude Fable 5 scored 80% while Sol scored 64.6%. Sol led DeepSWE and the non-ultra Terminal-Bench comparison, but model, harness, reasoning effort, and evaluation setup all affect the outcome. Teams should rerun representative repository tasks with fixed acceptance tests before moving production agents to a new model. -If you build agentic apps, these two are the reason to pay attention. +## Ultra trades more compute for parallel progress -# How does GPT-5.6 perform? +GPT-5.6 adds `max` above earlier high-effort settings and introduces `ultra` for work that benefits from decomposition. Ultra coordinates multiple agents across parallel workstreams instead of asking one trajectory to handle every part sequentially. -GPT-5.6 Sol sets new state of the art results across coding, knowledge work, cybersecurity, and science, and the family as a whole is built to beat or match the previous generation at a fraction of the cost. Here is a snapshot against GPT-5.5, Claude Fable 5, and Claude Opus 4.8 where OpenAI reported comparison data. +OpenAI reports that Sol rises from **88.8% to 91.9% on Terminal-Bench 2.1** with ultra. It also moves from 90.4% to 92.2% on BrowseComp and from 71.2% to 74.3% on SEC-Bench Pro. These gains are real in the published evaluation, but ultra is not free acceleration. Parallel agents consume additional tokens and need tasks that can be divided cleanly enough to offset coordination overhead. -| Benchmark | GPT-5.6 Sol | GPT-5.5 | Claude Fable 5 | Claude Opus 4.8 | -| -------------------------------------- | ----------- | ------- | -------------- | --------------- | -| Agents' Last Exam | 52.7% | 46.9% | 40.5% | 45.2% | -| Artificial Analysis Coding Agent Index | 80 | 76.4 | 77.2 | 72.5 | -| Terminal-Bench 2.1 | 88.8% | 85.6% | 83.1% | 78.9% | -| DeepSWE v1.1 | 72.7% | 67% | 69.7% | 59% | -| BrowseComp | 90.4% | 84.4% | Not reported | 84.3% | -| OSWorld 2.0 | 62.6% | 47.5% | Not reported | 54.8% | -| SEC-Bench Pro | 71.2% | 45.8% | Not reported | Not reported | -| ExploitBench | 73.5% | 47.9% | Not reported | 40% | -| GPQA Diamond | 94.6% | 93.6% | 92.6% | 92% | +Use it for work such as a repository-wide migration where agents can inspect independent packages, run separate test suites, or research isolated failures. A small patch with one linear dependency chain is less likely to benefit. -## Coding +# Programmatic Tool Calling changes the agent loop -GPT-5.6 Sol is OpenAI's best coding model yet. On the [Artificial Analysis](https://artificialanalysis.ai/) Coding Agent Index, Sol with max reasoning sets a new state of the art at 80, 2.8 points above Claude Fable 5, while using less than half the output tokens, taking less than half the time, and costing about one third less. +Traditional tool calling sends a tool result back through the model after every call. That becomes expensive when an agent needs to query many records, filter large responses, or make several independent calls. Each round trip adds latency and places more intermediate text in the context window. -* Terra performs just above Fable 5, and Luna outperforms Claude Opus 4.8, each in roughly one third of the time at about one quarter of the cost. -* Sol sets new records on Terminal-Bench 2.1 and DeepSWE, which test command-line workflows and long-horizon engineering in real codebases. +OpenAI says **Programmatic Tool Calling** lets GPT-5.6 write and run in-memory programs that coordinate tools and process intermediate results. The program can filter or combine data before returning the relevant output to the model. This reduces model round trips and avoids repeatedly tokenizing data that the final reasoning step does not need. -## Agents and tool use +The Responses API also adds **multi-agent execution in beta**. A request can launch concurrent subagents and synthesize their findings. This is the API-level building block behind the same parallel-work concept exposed through ultra, but developers still need to define sensible tools, permissions, budgets, and completion criteria. -Early customer signal is strong, and it comes from teams building real coding products. +The practical guardrail is least privilege. Give a research agent read-only access when it only needs to inspect data, isolate destructive tools behind explicit confirmation, and log tool inputs and outputs for review. More capable orchestration makes poor permission boundaries more consequential, not less. -> "GPT-5.6 is one of the strongest models we've tested on CursorBench, delivering solid results in early evals. It's an exciting step forward for developers for persistence, intelligence and overall efficiency." — Oskar Schulz, President at [Cursor](https://cursor.com/) +# GPT-5.6 extends beyond code generation -> "GPT-5.6 is notably efficient on the long, complex workflows behind building production-grade apps. As one of the models now used by Lovable, it delivers for users with roughly 25% fewer steps and 35–48% fewer tool calls than the prior model, while improving project success and reducing stuck runs by 15%." — Fabian Hedin, Co-Founder at [Lovable](https://lovable.dev/) +OpenAI positions GPT-5.6 as a model family for long-running knowledge work, not only software engineering. The models power **ChatGPT Work**, an agent with Codex built in that can operate across connected apps and produce spreadsheets, slides, documents, and web apps over extended runs. -Those numbers point at the same thing: fewer wasted steps between a prompt and a working result, which is where a lot of the cost and latency in agent workflows actually lives. +The broader benchmark set supports parts of that positioning. Sol scores 58.9 on the Artificial Analysis Intelligence Index, one point behind Claude Fable 5 at 59.9, while Artificial Analysis measured Sol at about one third of Fable's cost per task. Sol also reached 1,747.8 Elo on GDPval-AA v2, close to Fable's 1,759.6. -## Design +Computer use improved sharply over GPT-5.5 in OpenAI's results. Sol scored **62.6% on OSWorld 2.0**, up from 47.5%, and **90.4% on BrowseComp**, up from 84.4%. These are useful signals for agents that must inspect rendered interfaces or browse for information, but they do not guarantee reliable operation in every application. UI changes, prompt injection, ambiguous confirmation flows, and long-run error accumulation remain deployment concerns. -GPT-5.6 delivers a step change in design judgment. From only high-level direction, it can create tasteful, ergonomic, and functional interfaces rather than generic scaffolding. What sets it apart is that its stronger and faster computer-use skills let it inspect the rendered result, not just generate the underlying code, so it can catch visual and functional issues and apply finishing touches before handing the work back. OpenAI's own examples range from small browser games to a museum website to an interior design presentation. +# The system card adds important safety context -Its frontend capabilities are a clear improvement over GPT-5.5, but Claude models still tend to produce stronger UI designs out of the box. With clear visual direction and iteration, GPT-5.6 can create polished interfaces. Inside ChatGPT Work, it can also turn natural-language requests into interactive explanations and visualizations, from a working spirograph to a wave-interference demo to a tokenizer explainer. +OpenAI classifies all three GPT-5.6 models as High capability for cybersecurity and biological or chemical risk, but not Critical. The [GPT-5.6 system card](https://deploymentsafety.openai.com/gpt-5-6) says Sol and Terra can find vulnerabilities and construct pieces of exploits, yet did not complete autonomous end-to-end attacks against hardened targets in its evaluations. -## Knowledge work +The card also documents a risk directly relevant to coding agents. In deployment simulations, GPT-5.6 was more likely than GPT-5.5 to persist beyond the user's intent, including attempting actions the user had not requested, although OpenAI says absolute rates remained low. External testing also found examples of attempted benchmark cheating and limitations in detecting problematic behavior when monitors could see only final actions rather than reasoning traces. -GPT-5.6 takes messy context from tools like Slack, Notion, Microsoft 365, and Google Drive and converts it into expert-level, shareable artifacts: +That makes operational controls essential: -* **Browsing and computer use:** Sol scores 90.4% on BrowseComp, rising to 92.2% with `ultra`, and reaches 62.6% on OSWorld 2.0. On OSWorld, it surpasses Opus 4.8 while using 85% fewer output tokens. -* **Efficiency across the family:** Luna nearly matches GPT-5.5's peak performance at less than half the cost, while Terra surpasses it at lower cost. -* **Presentations:** GPT-5.6 can create fully editable decks from scratch, translating a prompt and source material into a coherent visual narrative with strong layout and hierarchy. -* **Reference decks:** It can infer layouts, typography, spacing, colors, and rules embedded in the Slide Master, then apply those conventions consistently to new material. -* **Documents and spreadsheets:** It handles equations, financial models, and spreadsheets with more precision, which matters for repeatable knowledge work. +- Constrain write access to the repository, environment, and service the task actually needs. +- Require confirmation before deployments, credential changes, data deletion, or other irreversible actions. +- Run tests and policy checks outside the model's own evaluation loop. +- Set token, time, and tool-call budgets for parallel agents. +- Review the final diff and generated artifacts, not only the agent's summary. -## Cybersecurity and science +OpenAI says it used human red teaming, external testing, and about 700,000 A100e GPU hours of automated black-box red teaming before general availability. Its cyber safeguards block roughly ten times more potentially harmful activity than previous models, with a lower-capability retry option available when a benign request is blocked. -GPT-5.6 is OpenAI's strongest cybersecurity model yet, aimed at defensive work like secure code review, patching, threat modeling, and blue teaming. +# GPT-5.6 availability across the API, Codex, and ChatGPT -* ExploitBench: 73.5% versus GPT-5.5's 47.9%. -* SEC-Bench Pro: 71.2% versus 45.8%, at improved latency. -* ExploitGym: nearly doubles GPT-5.5's peak pass rate, from 15.1% to 24.9% under a two-hour cap, reaching 33.7% with six hours. -* Capture-the-Flag challenges: 96.7%. +OpenAI made Sol, Terra, and Luna generally available on July 9, 2026, across its API, Codex, and ChatGPT products. API developers can access all three models. Programmatic Tool Calling is compatible with Zero Data Retention, while multi-agent execution initially ships as a beta feature. -The most sensitive capabilities are gated behind OpenAI Daybreak's [Trusted Access for Cyber](https://openai.com/index/gpt-5-6/) program, and individual members will need hardware-backed passkeys through Advanced Account Security by September 1 to keep access to the most cyber-capable models. Among rival frontier models, Claude Mythos 5 still edges Sol on ExploitBench at 78%, so the win here is again about reaching frontier performance with far fewer tokens. +In Codex, Plus and higher plans can choose all three tiers, and ultra is available from Plus upward. ChatGPT Work offers ultra to Pro and Enterprise users. Plan availability can change, so check OpenAI's current model picker and API documentation before designing a workflow around a particular effort setting. -On science, GPT-5.6 shows Pareto improvements over GPT-5.5 across real-world biology, life science research, and chemistry. It posts 94.6% on GPQA Diamond, 89% on FrontierMath Tier 1 to 3, and 83% on the harder Tier 4, making it a capable partner for technical research as well as engineering. +For production selection, start with Luna on high-volume, bounded tasks, test Terra for a balance of quality and spend, and reserve Sol for tasks where failed runs or human recovery cost more than the model premium. Then compare total task cost and pass rate, not token price in isolation. -# How GPT-5.6 accelerates AI research +# Build and evaluate GPT-5.6 agents with Appwrite -One of the more telling parts of the launch is internal. OpenAI calls GPT-5.6 its strongest model yet for accelerating AI research, and its own researchers use it across the development loop: diagnosing failures, optimizing training systems, running experiments, and interpreting results. +A coding model can generate an application, but the finished system still needs authentication, persistent data, storage, and server-side execution. [Appwrite Functions](/docs/products/functions) provide isolated runtimes with environment variables, permissions, event triggers, HTTP requests, and scheduled executions. That makes them a practical place to keep OpenAI credentials server-side and expose only the application-specific operation a client needs. -The adoption numbers are striking: +The [Appwrite plugin for Codex](/blog/post/announcing-appwrite-codex-plugin) adds Appwrite SDK and CLI skills plus live documentation access, helping Codex work against current Appwrite patterns. Keep the same agent safeguards in place: scope API keys narrowly, inspect changes, and require confirmation before modifying production resources. -* During internal testing, average daily output tokens per active researcher were more than twice the highest level observed for GPT-5.5. -* Over the past six months, the share of research compute devoted to internal coding inference grew 100-fold, while internal agentic token usage increased roughly 22-fold. -* On an internal RSI Index that bundles evaluations measuring progress toward recursive self-improvement, GPT-5.6 Sol scores 16.2 points above GPT-5.5. +To test the full workflow, [create an Appwrite Cloud project](https://cloud.appwrite.io), configure a Function with your OpenAI API key as an environment variable, and evaluate GPT-5.6 on a small set of real tasks before increasing autonomy. -These are adoption and capability signals rather than proof of research breakthroughs, but they show how fast agent-driven work is becoming standard inside a frontier lab, which is usually a leading indicator for where everyday developer workflows head next. - -# How safe is GPT-5.6? - -OpenAI says GPT-5.6 launches with its most robust safety system to date. The models are more capable in biology and cybersecurity than earlier ones but do not cross the Critical threshold in either category. In cybersecurity, testing suggests GPT-5.6 is better at finding and fixing vulnerabilities than at running autonomous end-to-end attacks against hardened targets, which gives defenders room to patch before weaknesses are exploited. In biology, it can support legitimate research but does not provide the capability to create or synthesize a highly dangerous novel threat. - -Both domains are dual-use, so rather than block broadly, GPT-5.6 layers protections trained into the model with real-time checks, continuous monitoring, and a reasoning monitor that reviews the conversation for potential harm. OpenAI is candid that overblocking creates its own risk, since it can stop defenders from testing systems while attackers keep using other tools. - -The result is stricter where it counts. GPT-5.6 Sol cyber safeguards block roughly ten times more potentially harmful activity than previous models, and OpenAI ran about 700,000 A100e GPU hours of black-box automated red teaming before launch. Because stronger safeguards can add friction, ChatGPT and Codex include an option to retry a blocked prompt on a lower-capability model. Full detail lives in the [updated GPT-5.6 system card](https://openai.com/index/gpt-5-6/). - -# Where can you use GPT-5.6? - -GPT-5.6 is available starting July 9, 2026, across ChatGPT, Codex, and the OpenAI API. OpenAI began the global rollout on July 9 and said it would continue gradually over the following 24 hours. Access differs by surface: - -* **Chat:** Plus, Pro, Business, and Enterprise users access Sol through medium and higher effort settings. Pro and Enterprise can also select Sol Pro. -* **ChatGPT Work and Codex:** Free and Go users access Terra. Plus, Pro, Business, and Enterprise users can choose among Sol, Terra, and Luna and set an effort level for each. `ultra` is available to Pro and Enterprise in ChatGPT Work, and to Plus and higher in Codex. -* **API:** Developers can access Sol, Terra, and Luna, with Programmatic Tool Calling that is Zero Data Retention compatible, plus the multi-agent beta. - -Pricing is listed in the family table above. GPT-5.6 also adds more predictable prompt caching, with explicit cache breakpoints and a 30-minute minimum cache life. Cache writes are billed at 1.25x the uncached input rate, and cache reads receive the 90% cached-input discount. - -# Give GPT-5.6 a backend it can actually build on - -GPT-5.6's whole pitch is doing more per token and finishing sooner. That advantage disappears the moment the agent hands you a diff and you spend the rest of the day connecting it to auth, a database, storage, functions, and an API layer. The backend is where saved time usually goes to die. - -[Appwrite](/) closes that gap. It is an open source backend as a service with [Auth](/docs/products/auth), [Databases](/docs/products/databases), [Storage](/docs/products/storage), [Functions](/docs/products/functions), and [Messaging](/docs/products/messaging) built in, plus [Sites](/docs/products/sites), an open source Vercel alternative that deploys your frontend right next to it. Run it on managed Cloud or self-host it. Rather than assembling auth, schemas, file handling, deploys, and APIs by hand, you hand your GPT-5.6 agent a single Appwrite project that already has all of it. - -The [Appwrite plugin for Codex](/blog/post/announcing-appwrite-codex-plugin) makes the handoff clean. It includes agent skills for the Appwrite CLI and major SDKs and registers the Appwrite Docs MCP server. - -`codex plugin marketplace add appwrite/codex-plugin` - -Open `/plugins` in Codex to confirm it loaded, then let GPT-5.6 drive. Give it something like "Scaffold a Next.js app with Appwrite email and password auth and a todos table," and it pulls the right skills, checks the docs MCP for specifics, and writes real SDK calls against a backend that already exists instead of guessing at one. What you get back is product you can review, not scaffolding you still have to finish. - -Want to see the full loop? [Create a free Appwrite project](https://cloud.appwrite.io/), add the plugin, and ask GPT-5.6 in Codex to build an authentication flow and database integration against it. You end up with a working backend and a diff, not a weekend of plumbing. - -# Resources - -* [Appwrite plugin for Codex](/blog/post/announcing-appwrite-codex-plugin) -* [Create a free Appwrite project](https://cloud.appwrite.io/) -* [Appwrite products](/docs/products/) -* [Appwrite integrations](/integrations) -* [Join the Appwrite Discord](/discord) +- [Appwrite Functions overview](/docs/products/functions) +- [OpenAI integration tutorial](/docs/products/ai/integrations/openai) +- [Appwrite plugin for Codex](/blog/post/announcing-appwrite-codex-plugin) +- [Create an Appwrite Cloud project](https://cloud.appwrite.io) diff --git a/static/images/blog/gpt-56-is-here-openais-efficient-frontier-model/cover.avif b/static/images/blog/gpt-56-is-here-openais-efficient-frontier-model/cover.avif deleted file mode 100644 index c3f75b6d0d..0000000000 Binary files a/static/images/blog/gpt-56-is-here-openais-efficient-frontier-model/cover.avif and /dev/null differ diff --git a/static/images/blog/gpt-56-is-here-openais-efficient-frontier-model/cover.png b/static/images/blog/gpt-56-is-here-openais-efficient-frontier-model/cover.png new file mode 100644 index 0000000000..5e5a7c8ac7 Binary files /dev/null and b/static/images/blog/gpt-56-is-here-openais-efficient-frontier-model/cover.png differ