GPT-5.6 のレビュー（1 分で読めます）

GPT-5.6 is finally here. Government approval held it up for two weeks,

and it lands head-on against Claude Fable. I've used GPT-5.6 Sol

internally for the past two months and probably burned through more than

25 billion tokens. I know what it's great at now. I also know where it

falls apart. Codex has gotten

much faster lately. It's very noticeable. And with Cerebras expected to

soon serve full GPT-5.6, not a watered-down version, we're going to see

frontier-model speeds unlike anything else we've seen. Speedmaxxis

rejoice! But raw tokens per second are not the only thing that makes it

fast. GPT-5.6 usually takes the short route to task completion. It knows

how to do things really well and directly. The Minecraft-style game is

the best example. I started it with

/goal and walked away. Once the core game worked, I expected

it to stop. It didn't. The model kept adding levels, mobs, and tiny

details. Days later, I killed the run myself. I never felt like I needed

to jump back in and steer it onto the right path.

That only happened because I used /goal. In that mode,

GPT-5.6 will work for an almost absurd amount of time without checking

in. With a normal prompt, it stops sooner and is more likely to ask what

comes next. Any claim about its long-run autonomy needs that caveat.

![RUN NOTES · LONG RUN

The run kept finding more work.

• 01

/goal started

One open build

• 02

Core game worked

Where I thought it would stop

• 03

Levels and bosses appeared

It decided the job wasn't done

• 04

Stopped manually

Days into the run

This level of persistence only showed up with /goal.](https://signals.forwardfuture.com/gpt-5-6-review/demo-screenshots/jade-ravine-ey5b.jpg)

Blockcraft showed it could keep going. The Excel clone was a harder

test: it had to inspect Excel on my desktop, recreate the interface,

and make a lot of interlocking pieces work together.

Try the Excel clone.

GPT-5.6 built this over six days. It's the real app, not a

screenshot. Edit a cell, write a formula, and poke around.

I gave it a finish line and let it run. It used Computer Use to inspect

Excel on my desktop, compare that against its own build, and keep closing

the gaps. I stopped the run myself after about six days. It is not

feature-complete, but it is a lot more than a demo. Give it enough time

and a clear target, and GPT-5.6 keeps finding the next thing worth

fixing.

The reasoning setting matters.

Light is good for quick questions and small local edits. High and Extra

High are the sweet spot for serious work. Ultra usually takes longer than

the extra thinking is worth and burns tokens. Fable is often good enough

at Medium or High. With GPT-5.6 Sol, I keep bumping it to Extra High.

I love that GPT-5.6 comes in three sizes. I can choose Sol, Terra, or Luna

for speed and cost, then set the thinking effort separately for each one.

Sol is the one for hard work. Terra is the everyday model. Luna is fast

and cheap. Those controls are precise. I just wish Codex routed my prompts

automatically.

It opens the browser first.

GPT-5.6 loves the Codex browser. If a task lives on a website, it usually

opens the site instead of asking for an API key or forcing the job into

the terminal. I switched back to GPT-5.5 to check. It went straight to

the command line, even when the browser made more sense.

Its browser control is the fastest and most accurate I've used. While a

data import ran, I told it to watch Supabase and resize the instances as

the load changed. It stayed on the dashboard and handled it. I didn't

set up an API or hand it keys. It used Supabase the way I would.

I gave GPT-5.6 a Google Workspace job I would normally have clicked

through myself: move Forward Future from

forwardfuture.ai to forwardfuture.com without

breaking the old addresses. It worked through the admin console the way

I would have. It verified the new domain, made .com primary,

kept the .ai aliases working, and set up MX, SPF, and DKIM.

It also paused before a consequential save, told me exactly what would

change, and waited for my confirmation.

Atlas

was OpenAI's AI-native browser: ChatGPT was built into the page, with

browser memory and an agent mode that could open tabs and take actions

for you. The idea was to get past the copy-paste-and-tab-switching

routine. It never really took off as the browser people lived in every

day. GPT-5.6 with Codex is the first version of that idea I want to use

all day.

The annoying work is rarely one big assignment. It's fifty little jobs:

check a dashboard, pull a number from one service, change a setting in

another, confirm the result, and get back to what I was doing. Codex is

unusually good at that loop. It opens the page, reads what is there,

takes the action, and checks that it stuck instead of asking me to build

an API integration or a one-off script first.

I am not ready to make it my actual main driver browser yet. I still

watch anything that can change production data or access. But the shape

of the workflow is real now: I browse normally, and when a chore appears

I hand it off in the same window. If that keeps getting more reliable,

the browser stops being where I do every click and becomes where I direct

the work.

The personality is bland. I like it more than I expected. GPT-5.6 gives

me enough information, explains things simply, and rarely pads the

answer. Claude still feels warmer. GPT-5.6 is more clinical, which makes

it easy to work with. I usually have to ask Fable to explain things more

simply or be more concise.

Its front-end taste has improved, but the default is predictable: every

site turns into a PowerPoint deck with giant statements and hard section

breaks. The work looks good. It also looks like slides. The model takes

design direction well, though. I kept pushing this page toward an essay,

and it followed without breaking the parts I liked.

![Design direction · before / after

Same product, different instincts.

The first pass looked like a PowerPoint deck. One round of

direction made it denser, quieter, and easier to use.](https://signals.forwardfuture.com/gpt-5-6-review/design-comparison/gpt-5-6-design-before-large.gif)

I still don't trust it blindly. I asked it to rebuild parts of my system

that had used OpenClaw. It said the job was done. Later, I discovered it

wasn't. You have to define "done" and give the model something it can

check. When it gets confused, bits of its internal process sometimes leak

into the answer. It's rare, but obvious when it happens.

The comparison, so far.

Fable made me want to invent a huge project so I could use it properly.

GPT-5.6 is easier to reach for. I use it for one-line edits, browser

chores, and big builds without changing how I work. GPT-5.6 feels about

2 to 3 times faster to me. That's an impression, not a benchmark. It

generates faster, wanders less, and tends to touch less code. GPT-5.6 is

an incredible model, easily the best of the GPT-5 line.

Let's start with the analogy, because everyone seems to be giving theirs.

GPT-5.6 is likely the last model in the GPT-5 training-run series. It's

an athlete at the peak of a long career.

Years of experience have made it the most reliable player on the court,

with the highest game IQ. But there's no more room to grow; this training

run has reached its ceiling.

Fable, by contrast, is essentially the first model from a new training

run: the first-round draft pick rookie. It has raw talent and the energy

of a young player, capable of incredible plays we didn't think possible.

It also makes mistakes that come with inexperience. Give that rookie more

games, and it may eventually be better than the veteran ever was. It's a

new game, and a new era.

The biggest differences show up before you ever look at a benchmark.

This table is the practical version: how each model takes direction, how

much latitude it gives itself, and how it moves through a task. It is

less scientific than the benchmark table below, but it has been more

useful for deciding what I use every day.

Personality

Blunt and a little bland. It usually gives me just enough

explanation.

Warmer. More natural to talk to.

What I give it

A small edit, a browser chore, or a large build. It works at any

scale.

I kept looking for a project big enough to justify using it.

Path to completion

Usually takes the direct route and stays inside the requested

scope.

More likely to expand the job when the prompt leaves room.

Autonomy

It can run for days with /goal. With a normal prompt, it

treats the job as bounded.

More naturally autonomous on large, open projects.

Perceived speed

Feels about 2 to 3 times faster. It generates quickly, takes fewer

detours, and changes less code.

Slower to generate and slower to work through the project.

The two launch pages publish results for 15 named benchmarks. Only

Terminal-Bench 2.1 is a clean head-to-head comparison: OpenAI reports

both GPT-5.6 Sol and Claude Fable 5 in the same chart. OpenAI publishes

GeneBench, ExploitBench, and ExploitGym as performance curves, so the

table shows the full Sol range from low through max reasoning. For most

of Anthropic's results, the source table combines Mythos 5 and Fable 5

and displays the higher score. Those values are included below, but are

labeled as combined rather than Fable-only results. Where the OpenAI

launch chart has no matching result, I also checked its linked system

card and the benchmark publishers' primary leaderboards.

Terminal-Bench 2.1

88.8%Sol · max reasoning91.9% in Sol UltraOpenAI chart ↗

84.3%Fable-only · same OpenAI chartOpenAI chart ↗Anthropic separately reports 88.0%* for its combined Mythos/Fable column.

GeneBench v1

10.8–30.7%Low → max reasoning · 2.7K–30.1K output tokensOpenAI chart ↗

Not published

ExploitBench Cap%

28–73%Low → max reasoning · 28.7K–120.5K output tokensOpenAI chart ↗

78.0%*Anthropic combined Mythos 5 / Fable 5 result · not Fable-onlyAnthropic table ↗

ExploitGym

6–25% 2h6–34% 6hLow → max reasoning · intended exploit rateOpenAI chart ↗

Not published

SWE-Bench Pro

No GPT-5.6 result found

80.3%Anthropic combined Mythos 5 / Fable 5 result · not Fable-onlyAnthropic table ↗

FrontierCode Diamond

No GPT-5.6 result found

29.3%xhigh · Anthropic combined Mythos 5 / Fable 5 result · not Fable-onlyAnthropic table ↗

GDPval-AA

No GPT-5.6 result found

1932Anthropic combined Mythos 5 / Fable 5 result · not Fable-onlyAnthropic table ↗

GDP.pdf

No GPT-5.6 result found

29.8%No tools · Anthropic combined Mythos 5 / Fable 5 result · not Fable-onlyAnthropic table ↗

Blueprint-Bench 2

No GPT-5.6 result found

38.6%Anthropic combined Mythos 5 / Fable 5 result · not Fable-onlyAnthropic table ↗

AutomationBench

No GPT-5.6 result found

17.4%Anthropic combined Mythos 5 / Fable 5 result · not Fable-onlyAnthropic table ↗

OSWorld-Verified

No GPT-5.6 result found

85.0%Anthropic combined Mythos 5 / Fable 5 result · not Fable-onlyAnthropic table ↗

Legal Agent Benchmark

No GPT-5.6 result found

13.3%Anthropic combined Mythos 5 / Fable 5 result · not Fable-onlyAnthropic table ↗

Humanity's Last Exam

No GPT-5.6 result found

59.0%* no tools64.5%* with toolsAnthropic combined Mythos 5 / Fable 5 results · not Fable-onlyAnthropic table ↗

BioMysteryBench

No GPT-5.6 result found

46.1%* hard83.9%* human-solvedAnthropic combined Mythos 5 / Fable 5 results · not Fable-onlyAnthropic table ↗

HealthBench Professional

60.5%Length-adjusted · 64.1% unadjustedOpenAI system card ↗

66.0%*Anthropic combined Mythos 5 / Fable 5 result · not Fable-onlyAnthropic table ↗

How to read Anthropic's numbers. Anthropic says Mythos 5

and Fable 5 are usually within 1–3 percentage points, so its launch table

shows the higher result. An asterisk marks benchmarks where their

safeguards create a larger gap; on those tests, Anthropic says Fable

performs closer to Opus 4.8 because some requests fall back to that

model. “No GPT-5.6 result found” means no primary-source result appeared

in the launch post, linked system card, or benchmark leaderboards I

checked—not a score of zero.

If Fable matched GPT-5.6 in the browser and on front-end work for a lot

less money, that would matter. If GPT-5.6 posted a higher benchmark score

but needed more retries to finish real work, that would matter too. The

number I care about is cost per finished task, not cost per token.

What the APIs cost. Standard API prices in USD per million tokens.

Input tokens

$5 / million tokens

$10 / million tokens

Cached input

$0.50 / million cache hits**$6.25 / million cache writes

$1 / million cache hits

$12.50 / million 5-minute writes

$20 / million 1-hour writes

Output tokens**

$30 / million tokens

$50 / million tokens

GPT-5.6 pricing and tiers: OpenAI ↗

· Fable pricing: Anthropic ↗

Three more things it built.

I keep using it.

GPT-5.6 wastes less of my time.

I keep reaching for it because I get to the result faster and spend less

time fighting the model. It isn't the best at everything, and

/goal changes how it behaves on long runs. For daily work, I

care about that more than a benchmark. It is the model I have the most

confidence in to get the job done right the first time.

Define "done." Give it something concrete to check. If the work lives on

a website, let it use the browser instead of shoving the job into the

terminal. It still needs supervision. A vague prompt can end with a

confident "done" before the work is done.

It is great to finally have a frontier OpenAI model in the mix while

Fable is out. GPT-5.6 is probably the last GPT-5-family model before

GPT-6, and it feels like the most complete version of that family: Codex

is more mature than Claude Code in the places I care about, especially

browser control and computer use. The token efficiency versus Fable is

absurd, and OpenAI has been generous with quotas. They have also been a

pleasure to work with, and unusually transparent about what was happening

during the government delay.