Is GPT-6 called Astra?

The Naming Confusion Around GPT-6

You've seen three different names for the same model this week: one headline says GPT-6, a code snippet says gpt-6, and the launch coverage keeps saying Astra. Here's the short answer: yes, GPT-6 is branded Astra, and the API model ID is gpt-6-astra, confirmed by OpenAI's official model documentation and launch coverage of the September 3, 2026 release.

The confusion mostly comes from codenames that leaked before the product shipped. Per one detailed writeup, "Astra" and "mewfour" both circulated as internal codenames, while the pretraining project itself was reportedly called "Doug." So if you saw any of those terms in rumors or changelogs months ago, they were all pointing at the same model that now carries the official Astra branding.

There's also a practical developer trap here. Older code snippets that use model="gpt-6" predate the launch naming and won't match the current API. The correct ID is gpt-6-astra, and if you're wiring it into an agent, that distinction matters the moment your tool call fails.

The GPT-6 family follows a celestial naming scheme that helps you pick the right tier:

• Astra: the flagship for complex reasoning and coding, at $10 per million input tokens and $50 per million output.
• Sol (GPT-6.1): near-Astra quality at roughly 5x lower cost, $2/$10 per MTok.
• Luna: the efficient tier at $0.10/$0.50 per MTok, built for high-volume work.

All three share a 1.05M-token context window and 128K max output, per the official docs.

One error keeps spreading, and it's worth correcting directly: people keep calling GPT-5.6's flagship "Astra." It isn't. Sol is GPT-5.6's flagship tier, so "GPT-5.6 Astra" is simply a naming mistake, per the same source. The timeline makes the mix-up understandable: GPT-6 arrived roughly two months after GPT-5.6, which itself came less than three months after GPT-5.5, as Zapier's model guide notes. OpenAI is shipping point-releases faster, and the names haven't always kept pace in people's heads.

If you're pointing Astra at agentic work, one more thing: researcher Sebastian Raschka has suggested that newer models can be over-constrained by stale AGENTS.md and SKILL.md instructions, so it's worth pruning old hand-holding content. If you're refreshing your skill setup, SkillzBench's free directory pins every skill to a verified commit and security-scans each one, and its free MCP connector lets you toggle exactly which skills each agent sees.

Why OpenAI Chose Celestial Names: Astra, Sol, and Luna

If you've been staring at OpenAI's model picker wondering why one family has three names and none of them are numbers, there's a reason. The GPT-6 generation uses a deliberate celestial naming system, confirmed in OpenAI's own developer documentation: Astra for the flagship, Sol for the mid-tier, and Luna for the efficient workhorse.

The structure maps cleanly onto price and capability. Astra is the flagship, the model OpenAI's docs point to first for complex reasoning and coding, priced at $10 per million input tokens. Sol, attached to the GPT-6.1 point release, runs at $2/$10 per MTok, roughly 5x cheaper while staying "near-Astra" in capability. Luna sits at $0.10/$0.50 per MTok, about 100x cheaper than Astra on input, built for high-volume work where cost matters more than frontier reasoning. All three share the same 1.05M-token context window and 128K max output, so the tier difference is capability and price, not capacity.

The names also follow a consistent technical pattern. Model IDs use kebab-case: gpt-6-astra, gpt-6.1-sol, gpt-6-luna. That matters practically, because older code snippets referencing plain model="gpt-6" predate the launch naming and won't resolve correctly.

Here's where it gets messy from the outside. Astra doesn't live in isolation; it sits in a broader OpenAI lineup that mixes numeric versions, celestial codenames, and domain labels. You'll find GPT-5.6 Cyber (security-focused), GPT-Rosalind, and image models like GPT-Image-2.5 Sunburst and Flare alongside the GPT-6 family. Zapier's hands-on review of the lineup joked about OpenAI's "crimes against consistent naming conventions, capitalization, and hyphenation," and it's hard to argue. One common error worth flagging: "GPT-5.6 Astra" doesn't exist. GPT-5.6's flagship tier is Sol, so that combination is a naming mix-up.

The inconsistency has a logic, though. OpenAI has said it's serious about pushing point releases faster rather than waiting long stretches between frontier models, and it's comfortable with different models carrying different version numbers. GPT-6 arrived about two months after GPT-5.6, which itself came less than three months after GPT-5.5. When you ship that fast, naming each release as a clean sequential version gets unwieldy, so the celestial suffixes act as a family brand that survives rapid iteration.

A practical note for agent builders: naming conventions shift, but the instruction packages your agents run on shift too. Sebastian Raschka points out that newer models can be over-constrained by stale AGENTS.md and SKILL.md content, so it's worth pruning what your agents carry. If you're auditing which skills actually help, the free directory and MCP connector at SkillzBench benchmark each skill against the same model with and without it, so you can see the lift before you load it into an agent.

What the Astra Name Signals About Capability and Cost

The name on the box tells you where a model sits in OpenAI's lineup, and "Astra" is the label reserved for the top of the stack. According to OpenAI's own developer documentation, the GPT-6 generation ships under celestial suffixes: Astra as the flagship, Sol as the mid-tier GPT-6.1 option, and Luna for efficient, high-volume work. So "GPT-5.6 Astra" is a naming error, since GPT-5.6's flagship tier is Sol.

What you actually get under that name is a model priced and specced for heavy autonomous work. OpenAI's docs list a 1.05M-token context window, 128K max output, a knowledge cutoff of April 30, 2026, and reasoning levels running from low up to max. Pricing lands at $10 per million input tokens and $50 per million output tokens, with a Fast mode at $20/$100 for roughly 2.5x speed.

Launch positioning leaned hard into computer use. OpenAI's September 3, 2026 rollout billed Astra as "the world's best computer use model," framed by president Greg Brockman with "Welcome to the AGI era," and gated first through the enterprise Daybreak program before reaching ChatGPT tiers, the API, AWS Bedrock, and Microsoft Azure.

The headline benchmark supports the pitch, with caveats:

• OSWorld 2.0 (offline subset): 72.6% in about 40 minutes per task, versus 65.7% for GPT-5.6 Sol at roughly 75 minutes.
• FrontierMath Tier 4 v2: 97.6%; GPQA Diamond: 96%; ARC-AGI-3: 98.6%.
• Independent coding results are described as mixed, and Artificial Analysis indexes place Astra at the frontier without a runaway lead.

That gap matters for anyone building agents. If your workflows depend on hand-holding instructions in AGENTS.md or SKILL.md files, Sebastian Raschka notes newer models may be over-constrained by stale instructions, worth pruning before you assume a capability jump will show up in your stack. SkillzBench's free directory and MCP connector let you verify and benchmark the skills you keep, so you're testing them against the model rather than trusting a launch slide.

The Developer Detail That Trips People Up: Correct Model IDs

If your API call fails with a "model not found" error on model="gpt-6", that's not a bug in your code. Older code snippets circulating online simply predate the launch naming, and the correct ID is gpt-6-astra, per developer documentation.

The GPT-6 family follows a celestial naming scheme, and the IDs matter when you're wiring up agents. OpenAI's official model docs list three tiers, all sharing a 1.05M token context window and 128K max output:

• gpt-6-astra: the flagship, at $10/M input and $50/M output
• gpt-6.1-sol: near-Astra quality at $2/$10, roughly 5x cheaper
• gpt-6-luna: the efficient tier at $0.10/$0.50, about 100x cheaper than Astra on input

That spread is the real design decision. Astra costs $10 per million input tokens; Luna costs a dime. If you're building an agent that runs hundreds of calls per task, routing the routine steps to Luna and reserving Astra for hard reasoning isn't optimization anymore, it's the difference between a viable product and a burned budget.

There's also a fast mode worth knowing about: $20/$100 per MTok buys you 2.5x speed at 2x price. For time-sensitive computer use, that trade can pay for itself. Cache pricing helps too, with reads at $1.00/M and writes at $12.50/M per published figures, so agents that repeat context benefit from caching.

The bigger signal is where pricing is heading. Brockman has hinted that OpenAI is moving toward price-per-task framing, saying "pricing tokens doesn't make any sense" per launch coverage. If per-task billing lands, model IDs become less of a cost lever and the quality of your agent's instructions and tooling becomes more of one.

That's worth acting on now. Researcher Sebastian Raschka, citing the Claude Code lead, notes newer models may be over-constrained by stale hand-holding instructions, so consider pruning old SKILL.md and AGENTS.md content. If you're rebuilding your agent's skill set, SkillzBench's free directory security-scans and benchmarks agent skills in a sandbox, publishing the measured Lift with and without each skill, and its free MCP connector lets you equip agents per loadout. Given that published research found roughly 26% of skills carry at least one vulnerability, verifying before you equip is cheap insurance.

What Astra's Arrival Means for Your Agent Skills

If Astra really is a step change in computer use, your agent's instruction files deserve a second look. Sebastian Raschka, echoing a point made by the Claude Code lead, notes that newer models can be over-constrained by old hand-holding instructions sitting in AGENTS.md and SKILL.md files. A model that was trained to need training wheels doesn't benefit from them.

That matters more with Astra than with past upgrades. OpenAI's launch materials position it as "the world's best computer use model," and it posted 72.6% on the OSWorld 2.0 offline subset while finishing tasks roughly 47% faster than GPT-5.6 Sol (per launch coverage). Zapier's hands-on testing found it a top performer on messy, multi-step workflows where accuracy matters most. Instructions written for a weaker model can actively hold back a stronger one, so it's worth pruning or archiving stale skill content before you judge the new model's output.

Here's a quick audit for your own setup:

• Open your AGENTS.md and SKILL.md files and flag anything that explains basics the model already handles, like "write tests before committing."
• Archive rather than delete, so you can restore guidance if a regression shows up.
• Re-test one real workflow with and without the pruned instructions and compare results directly.
• Check where each skill came from, since published research found about 26% of skills carry at least one vulnerability.

That last point is where most teams get burned. Skills are just instruction packages, and nobody vets most of what's floating around GitHub. This is the problem SkillzBench (skillzbench.com) was built for: it's a free, independent directory that continuously crawls and dedupes skills from GitHub, pins every version to a commit SHA, license-checks and security-scans each one, and publishes a trust score from 0 to 100. It also benchmarks every skill in a sandbox with the same model, with and without the skill installed, and shows the Lift in percentage points so you can see what a skill actually earns.

Once you've picked your skills, the free SkillzBench MCP connector and per-agent loadouts let you equip each agent with only the skills you choose. The agent sees nothing you didn't toggle on, which is a cleaner security posture than granting a blanket skill library. Everything on the platform, including the benchmarks and alerts when upstream skills change, is free.

Astra's arrival is a good excuse for this cleanup regardless of which model you run. Skills, proven: browse and benchmark the directory at skillzbench.com before your next agent upgrade.