AGI is still a marketing word: what Fable 5, Mythos 5, and GPT-5.6 actually measure

Every couple of months someone asks me, half-joking, whether the latest release is "basically AGI now." This month it's a fair question, because three things landed close together. Fable 5 came back from a fifteen-day government-ordered blackout and reclaimed the top SWE-bench slot. GLM-5.2 rushed out without a real benchmark to fill the gap Fable's absence left. GPT-5.6 shipped in three tiers with no SWE-bench Verified number at all, just a Terminal-Bench claim. None of that is evidence about general intelligence, not directly. But look at what each vendor chose to publish, and what they didn't: it tells you where the frontier actually sits right now, and it isn't where "AGI" would put it.

The number that moved, and the number that stopped moving

SWE-bench Verified (real GitHub issues, pass or fail against real tests) went from 62% (Sonnet 3.7, Feb 2025) to 95% (Fable 5, mid-2026) in about sixteen months. That's the steepest jump on this yardstick yet, and it's also, if you look at the full table, oddly isolated. Fable 5 sits at 95.0, Opus 4.8 at 78.0, GLM-5.2's provisional figure also lands at 78.0: a seventeen-point gap between the leader and everyone else on the same benchmark. One training run landed somewhere the rest of the field hasn't followed, rather than three labs converging on a ceiling, on a benchmark that's been hammered on hard enough across enough release cycles (curated tasks, real contamination risk) that a seventeen-point outlier is reason to check the eval before crowning a winner, not proof the frontier just leapt forward.

Either read, jump or plateau, lands in the same place: a bigger number on one fixed task set isn't the same claim as "general intelligence," and the gap between those two claims is most of the distance still left to cover.

GPT-5.6 skipped the number that matters, and that's the tell

Sol, Terra, and Luna, the GPT-5.6 family, shipped with a Terminal-Bench-leader claim and no published SWE-bench Verified score. Maybe it's a genuine strategic pivot toward a different eval. Maybe the number wasn't flattering and Terminal-Bench was. I don't know which, and neither do you, and that's the point: when a vendor gets to choose which yardstick reaches your screen, you're reading positioning, not capability. GLM-5.2 did the same thing from the other direction, shipping an "indicative" figure with a straight admission it hadn't been properly measured yet, because being first to fill Fable's empty chair mattered more that week than being right. A system genuinely closing in on general intelligence doesn't need its vendor to pick the friendliest test. Benchmark selection turning into a marketing decision is evidence the industry is optimizing for narrative this quarter, not evidence it's near a generality threshold.

The fifteen-day blackout was the real AGI test, and it quietly failed

The more interesting data point wasn't a benchmark at all. On June 12 the US government forced Fable 5 and Mythos 5 offline worldwide, overnight, on a national-security order. Fifteen days later the order lifted. In between: teams with provider abstraction and a validated local floor had an annoying two weeks. Teams without one had a worse two weeks. Nothing that would make the front page outside the AI trade press happened.

A model that can be switched off worldwide by one government's memo, and whose absence for two weeks barely registers past the AI trade press, is not what "general intelligence" was supposed to feel like.

If a system were doing genuinely general-purpose, economically load-bearing cognitive work at anything close to human-substitute scale, pulling it for two weeks should have been a macro event, not a shrug the market absorbed the way it absorbs a cloud region blinking. It wasn't. Fable is, by the numbers, the strongest model there is right now. This is a read on the category, not a knock on Fable specifically. We're still building very good, specialized tools you rent by the token, not something civilization has actually been restructured around yet. The blackout proved that by how little it broke.

What the advisor tool quietly admits

The advisor pattern (a cheap executor model that consults a stronger model mid-generation instead of running the expensive model end to end) is a good piece of engineering. It's also an admission. It exists because no model, at any price, is trusted to run long unsupervised without borrowing judgment from somewhere else. Multi-agent orchestration, eval gates, human-in-the-loop review, output guardrails: the scaffolding built around these models has gotten more elaborate every year, not less. If the base models were actually closing in on general intelligence, that harness should be shrinking as the core gets more trustworthy on its own. It's growing instead. Everyone building production agents already knows why: left alone for long enough, things fail in ways that are almost never about raw model capability, and it comes down to context loss, goal drift, and nobody having a model that can be trusted to notice its own mistake three hundred steps in.

What's actually still missing

Concretely, not hand-wavy:

• No persistent memory that isn't hand-built. A model doesn't learn from Tuesday's mistake on Wednesday unless someone engineers a pipeline to make that happen. The learning, if any, happens in your logs and your eval set, not in the model.
• No long-horizon autonomy anyone actually trusts. Every serious production agent I've seen gets supervised, capped, and rolled back constantly. That's a response to what happens without it, not caution for its own sake.
• No transfer to genuinely novel problems. Every number cited above, including the ones I just spent six paragraphs picking apart, is a score on a curated, pre-scoped, static task set. The moment a problem's boundaries aren't already drawn for it, today's systems degrade in ways a competent generalist wouldn't.
• No goals of its own. Everything discussed here still runs inside someone else's system prompt, executing someone else's intent. That's most of what the word "general" was supposed to mean in the first place, not a minor missing feature.

Where that leaves the estimate

I don't think AGI is close. What this cycle made obvious is that "AGI" and "state of the art" have quietly become two separate conversations that just happen to share a leaderboard. Fable's comeback, GLM's rushed provisional launch, GPT-5.6 skipping the benchmark that would've invited a direct comparison: none of it is a step toward general intelligence. It's competition on cost, context length, and how well the harness around the model compensates for what the model still can't do alone. That's a perfectly good industry, worth building a career on, worth the routing and cost architecture this whole site obsesses over. It's just not the AGI industry, whatever a fundraising deck says this week.

My honest guess: years, plural, and probably a different architecture than just a bigger version of whatever shipped this month, not more of the same curve. Check back after the next model that publishes a benchmark it didn't get to pick first.