GPT-5.6 Scores Higher in Benchmarks but Loses to Fable 5 in Real‑World Use

01. Official benchmarks and third‑party data

OpenAI announced the full rollout of GPT‑5.6 at 1 a.m. and presented internal benchmark graphs where GPT‑5.6 Sol scores lead Claude Mythos 5 by about four points on Terminal‑Bench 2.1 and surpass Fable 5 by more than ten points on the Agents' Last Exam. DeepSWE 1.1 shows a single‑task cost of roughly $8, less than half of Fable 5.

However, the author cautions that official graphs are biased, noting that every release event makes OpenAI look like the winner.

Vertical comparison on OpenAI's own AGI Index shows a jump from 43 % (GPT‑5.5) to 60 % for GPT‑5.6, a larger increase than previous version increments.

Horizontal comparison from the third‑party Artificial Analysis (Intelligence Index v4.1) ranks Fable 5 at 60 points (first) and GPT‑5.6 Sol at 59 points, a one‑point gap. Cost per Index task is $1.04 for GPT‑5.6 versus more than double for Fable 5. On the Coding Agent leaderboard, Codex + Sol scores 80 points, beating Claude Code + Fable 5's 77 points. Anthropic responded by resetting all users' 5‑hour and weekly token limits.

02. Personal hands‑on test

The author reset the Codex quota and gave a previously completed Fable 5 project to GPT‑5.6. While the task (a simple project) completed in just over an hour, GPT‑5.6 felt “rigid” and required frequent manual verification, whereas Fable 5 finished the same work in about ten minutes with perfect results.

For routine queries and loops, GPT‑5.6 performed quickly and reliably. Token consumption was a major issue: using a 20× Pro quota, two tasks consumed 80 % of the five‑hour limit, contradicting OpenAI’s claim of halved per‑task cost.

Conclusions from this single‑night test:

General tasks: GPT‑5.6 can compete effectively.

Deep project collaboration: still less smooth than Fable 5.

Token usage: drains far too quickly.

03. New information from the release

Codex is renamed “ChatGPT Work”, an agent‑focused workspace for non‑programmers with minor UI tweaks.

The desktop app now unifies Chat, Work, Codex, Scheduled, and Sites into a single application; the old app is renamed “ChatGPT Classic”.

GPT‑5.4 will be retired on July 23, ending its low‑price availability.

“Sites” feature lets users publish interactive webpages directly from chat, generating a shareable URL.

OpenAI reduced the frequency of the “goblin‑like” speech pattern from 0.405 % to 0.032 %.