GPT-6 Astra & Claude Fable 5.1:
benchmarks and the AGI debate.

What changed with Astra and Fable 5.1?

What these models accomplish also depends on their tools, instructions, memory, permissions, and review process. Choosing the model is one part of building the system.

You can improve how the system finds documents, checks a change, or asks for help. Those decisions can improve the result alongside any gain from a more capable model.

What the benchmark setup changes

Benchmarks let us compare systems using repeatable scoring rules. Check the task version and test conditions before comparing two numbers. Results can change with the setup, as the ARC-AGI example below shows.

Use the table to shortlist candidates. Then compare them on your own tasks, including the time you spend checking and repairing their output.

Which model belongs in your workflow?

Run a small pilot that compares the complete setup, including tools and human review. The steps below are a suggested protocol. This journal hasn’t run an Astra or Fable trial.

1. Define a pass before generating an answer.

   For a bug fix, require the relevant tests to pass, the reported behavior to be corrected, and the reviewer to accept the change. For a document answer, require the claim to be supported by the correct source. Keep critical failure rules separate from style preferences.

2. Record the setup.

   Keep task inputs, available documents, tool permissions, time limits, and retry budgets consistent. Record model IDs, dates, prompts, reasoning settings, and harness versions. If you can’t match a condition, name that difference in the result.

3. Keep unfamiliar cases aside.

   Use some cases to develop the workflow and different cases to evaluate it. Include missing information, ambiguous requests, and mistakes the system should catch. A small pilot can expose failures without establishing how reliably the system handles every possible task.

4. Count the work around the model.

   Log rejected attempts, retries, review time, and corrections. A fast first answer may still be expensive to turn into accepted work. Use the calculator below to make the cost of that work explicit.

5. Set limits on where you’ll use it.

   Check that the setup meets your quality requirements and permission boundaries before comparing cost and turnaround. Revisit the decision when the model, task, or workflow changes. A lower bill doesn’t compensate for an unacceptable error.

Are we entering the AGI era?

AGI usually describes broadly human-level cognitive capability across many tasks. The exact threshold is disputed. Google DeepMind’s From AGI to ASI uses a working definition around median human performance and distinguishes it from superintelligence that exceeds large, coordinated expert groups.

ARC Prize calls the result meaningful progress in generalization while explicitly declining to treat benchmark saturation as proof of AGI. Its environments have bounded rules and goals. The scoring methodology combines completed levels with action efficiency relative to human participants. “99.9%” is not a percentage of human intelligence.

Independent reporting from ACS Information Age captures the same disagreement. Researchers Toby Walsh and Rebecca Johnson question whether benchmark results justify broad claims about general intelligence.

What Altman, Amodei, and Brockman hope AI will make possible

The people building AI have ambitions that extend well beyond the next benchmark. Their essays explain those ambitions in their own words. Read them as arguments about the future, separate from the evaluations above.

Sam Altman: more people able to build

In Reflections, published in January 2025, Altman described a shift in ambition from AGI toward superintelligence. He argued that more capable tools could accelerate scientific discovery and increase prosperity, and connected gradual deployment with learning from use and giving society time to adapt.

His February 2025 essay Three Observations argues that cheaper access to useful intelligence could make more work feasible. He imagines software agents as coworkers that still need human direction and supervision. Greater prosperity, he cautions, doesn’t automatically create greater equality.

A developer can start with a smaller question: is a useful project now within reach? That might be a local-language support tool or a better way to prepare technical guidance.

Dario Amodei: faster scientific research

In his October 2024 essay Machines of Loving Grace, Amodei imagines AI exceeding leading experts across many fields and accelerating scientific and health advances. His scenario assumes that powerful AI exists first. Experiments, data, and physical constraints still affect what happens next. Today’s benchmark scores don’t establish that this scenario has arrived.

Greg Brockman: put discoveries to use

In The OpenAI Mission, co-written with Ilya Sutskever in March 2019, Brockman imagined systems that could help turn scientific breakthroughs into products and services, including more accessible healthcare. They made safety and the distribution of benefits central to that ambition. It’s a historical mission statement about what they wanted to build.

What could this mean for the IT industry?

A small team could try an improvement it previously postponed. Someone who knows a local industry could test a product idea. A student could attempt a more demanding project. These are practical reasons to be interested in better AI tools.

They don’t guarantee new jobs. Some tasks will shrink, and the transition won’t be even. Training, access, and how employers manage the change will affect who benefits.

If you’re deciding what to learn or build, consider these areas:

Junior developers still need to debug, model data, read unfamiliar code, and explain tradeoffs. Use an assistant during practice, then investigate and run the code it produces. Can you spot an incomplete answer and work out what’s missing?

Experienced engineers can support that learning by reviewing the reasoning behind a change, setting realistic exercises, and giving people responsibility gradually.

A 90-day project plan

Pick a manageable project and try this learning plan. The goal is to gather evidence about what works, with enough time to investigate what doesn’t.

1. DAYS 01–30

   Understand one workflow

   Choose a recurring task. Record how it works today, where errors occur, and who checks the outcome. Build a baseline with a handful of representative examples. Learn the underlying tools well enough to spot an incorrect result.

2. DAYS 31–60

   Build a small assistant you can review

   Connect only the data and tools it needs. Define the actions it can take and the points that need human review. Add tests, logs, and a straightforward way to undo a change. Compare model options on the same examples.

3. DAYS 61–90

   Test it and write up the results

   Test with real users and unfamiliar cases. Report failures as clearly as successes. Write up the problem, your decisions, evidence of improvement, and remaining limits for your portfolio. Explain what the results support, even if the improvement is modest.

Beyond AGI: ASI and the singularity

Artificial superintelligence, or ASI, concerns capabilities beyond human-level general intelligence. The singularity imagines technological change accelerating beyond our ability to predict it, potentially through AI improving AI. DeepMind’s research report explores possible routes and uncertainties.

Altman’s June 2025 essay The Gentle Singularity imagines AI-assisted research helping improve later AI systems. He distinguishes that feedback from a system autonomously rewriting itself. Whether that progress can sustain itself remains something to demonstrate.

More accessible learning, better research tools, and software for overlooked communities are goals worth working toward. Bring educators, designers, security specialists, and domain experts into that work alongside researchers and software engineers. Their knowledge helps decide what to build, who can use it, and how to check that it works.