This morning,
we just wrote about:
Claude Opus 5.5
significantly lowering the cost of an AI task.
Tonight,
OpenAI takes the same battle:
a step further.
On September 22 (US time), OpenAI officially released:
GPT-6 Sol
and
GPT-6 Luna.
If you simply think of them as:
"Two smaller models below GPT-6 Astra,"
you’d miss the real point.
Because OpenAI is focusing not on:
how large the model is,
but rather on:
how much real work can be done per dollar.
GPT-6 Officially Divides into Three Tiers
The top remains:
GPT-6 Astra.
OpenAI still positions Astra as:
the most capable model in the full GPT-6 family.
For the trickiest,
most important,
and least negotiable tasks,
Astra is still the go-to.
But the thing is:
Not every email,
document organization task,
or bug requires:
Astra.
So OpenAI added two lower tiers.
Middle Tier: GPT-6 Sol
Sol is positioned for:
complex daily work.
For example:
coding,
enterprise knowledge work,
computer use,
cross-app workflows,
and long-running agent tasks
that require reasoning but don't merit the cost of Astra every time.
You can think of it as:
the workhorse model.
Lower Tier: GPT-6 Luna
Luna emphasizes:
speed,
volume,
and cost-effectiveness.
For example:
document summarization,
data extraction,
classification,
short questions,
large-scale clerical work,
and clearly defined repetitive tasks.
In other words,
the GPT-6 family no longer just answers:
"Which is the strongest model?"
Instead, it answers:
"Which tier suits which type of work best?"
Immediate Change: Sol API Prices Halved
Let's look at prices.
For GPT-5.6 Sol:
Input tokens per million cost:
$4.
Output tokens:
$20.
For GPT-6 Sol:
Input:
$2.
Output:
$10.
That's a direct
50% reduction.
And not just cheaper input,
but even output—usually the more expensive part—was halved as well.
Luna Is Even More Affordable
GPT-5.6 Luna:
Input:
$0.20.
Output:
$1.20.
GPT-6 Luna:
Input:
$0.10.
Output:
$0.50.
This means a huge volume of simple tasks have dropped to very low token costs.
What does this mean?
Not just:
chatting becomes a bit cheaper.
It means many companies that
used to think "running this task tens of thousands of times a day with AI is too costly,"
might need to recalculate.
Why Can OpenAI Lower Prices Again?
OpenAI mentioned two key directions:
inference efficiency
and
caching.
That is,
it’s not simply discounting the model price.
The underlying serving infrastructure
is also optimized to handle
more work per unit of compute.
If the context already processed in a prompt
can be effectively reused,
there’s no need to recompute the entire prefix
for every request.
This aligns with what Claude Opus 5.5 discussed today about:
prompt caching,
reflecting the same ongoing battle.
AI Companies Realize Context Recalculation Is Too Costly
Agent tasks are getting longer.
An agent may:
first read company rules,
then scan documents,
open websites,
make edits,
check results,
and make revisions.
Every turn carries prior context.
If each step recalculates from scratch,
costs explode quickly.
So,
caching
is no longer a minor feature only API engineers care about.
It’s becoming
the core enabler for large-scale commercialization of AI agents.
OpenAI Shifts Benchmarking to Cost per Task
This is the most noteworthy today.
Previous model releases often focused on:
accuracy,
scores,
and pass rates.
Now OpenAI is telling you directly:
how much a completed task actually costs.
For example, they released:
a test where AI agents
work across:
sales, marketing, operations, customer support, finance,
and HR,
using
47 different tools to complete
end-to-end workflows.
GPT-6 Sol xhigh: Approx. $0.27 per Task
According to OpenAI,
GPT-6 Sol:
xhigh effort
scored
33.2%
on AutomationBench.
The calculated cost per task is roughly
$0.27.
In the same table,
GPT-6 Astra Low
scored
30.3%
but the cost per task was about
3.9 times higher than Sol.
This doesn’t mean
Sol is outright better than Astra,
because effort levels differ.
Benchmarks reflect
specific task sets.
Astra remains OpenAI’s high-end model
for the toughest tasks.
But this reveals a key product strategy:
not every task needs Astra.
OpenAI Even Compares Directly to Claude
Using public data from AutomationBench,
OpenAI compares:
Claude Opus 5 Max scored
26.9%,
with cost about
11.1 times
higher than GPT-6 Sol xhigh.
Claude Fable 5.1 combined with Opus 5 fallback scored
31.4%.
Its cost exceeds Sol’s by
about 8.9 times.
Caution is needed:
this is an
OpenAI-released benchmark.
Comparing competing models’ versions, effort levels, fallback strategies,
and pricing models can vary.
So this chart alone can’t prove:
GPT-6 Sol outright beats Claude.
The takeaway is that
OpenAI publicly shifts competitive focus from:
benchmark score
to
score divided by cost.
Anthropic Is Making the Same Move
One of Claude Opus 5.5’s biggest selling points today is:
typical workload costs reduced by about
40% compared to Opus 5.
Anthropic even warns:
don’t only look at
cost per million tokens,
but also at
cost per completed task.
A few hours later,
OpenAI includes
cost per task
at the core of the GPT-6 release.
Both leading AI firms
send the same signal almost simultaneously:
The next major AI model battle is an economics battle.
Coding Also Focuses on "Merge Ability & Cost"
OpenAI added a key new test:
FrontierCode.
It doesn’t just ask:
can the code run,
but checks:
tests,
scope discipline,
code style,
codebase standards,
and ultimately,
whether it’s close to
being mergeable.
Because enterprises with coding agents
fear not that AI can’t write code,
but that code looks correct yet requires serious revision by senior engineers.
If models are cheap but
engineers must spend an hour per task fixing code,
it’s still costly.
On DeepSWE Tests, Sol Nearly Matches Top Claude
In OpenAI’s DeepSWE 1.1 test,
GPT-6 Sol Max scored
68.8%,
while Claude Fable 5 xhigh scored
69.9%,
a difference of just
1.1 percentage points.
OpenAI estimates Sol’s cost per task is about
80%
lower.
Again, this is a comparison based on public competitor benchmarks,
not a universal statement that all coding workloads are 80% cheaper.
But it shows a crucial shift:
the assumption that "cheaper models only do simple work" is rapidly breaking down.
Luna Is Even More Interesting
GPT-6 Luna Max scored
66.6%
on the same DeepSWE test.
OpenAI says that’s roughly equivalent to
medium effort from Claude Opus 5 and Claude Fable 5,
but its cost per task is about 93% lower than Opus 5,
and 96% lower than Fable 5.
If third-party testing confirms these cost-performance gaps in real work,
the impact could be huge.
Many companies don’t need the
"world's strongest coding model,"
but instead
"one that can cheaply handle thousands of runs daily and is usually good enough."
Computer Use Enters the Price War
GPT-6 Astra
remains OpenAI’s strongest
computer use model.
However, in OSWorld 2.0 Offline
simulated computer operation tests,
GPT-6 Sol xhigh scored 60.5%,
Claude Opus 5 Medium scored 60.3%.
OpenAI claims Sol's per-task cost is roughly
80%
lower.
GPT-6 Luna Max surpasses
GPT-5.6 Sol Medium,
at only about one-tenth the cost.
This suggests that future computer use breakthroughs will be less about adding more app support,
and more about
enabling agents to operate hundreds of times a day affordably.
For AI Agents to Become Widespread, Solving "Too Expensive" Is Key
If an AI agent cleans up one email a day at a company,
almost nobody cares about cost.
But if it handles:
10,000 invoices daily,
5,000 leads,
2,000 support tickets,
hundreds of website flows,
a difference of just
10 cents per task can mean
thousands of dollars a day,
and hundreds of thousands a year.
So when AI transitions from
chatbot
to
digital worker,
price
suddenly becomes part of product capability.
Factuality Is Another Important Cost Factor
When the model makes mistakes,
it’s not just
a quality issue.
It’s also a cost issue,
because mistakes require humans to:
review,
ask again,
add data,
and rerun.
OpenAI specifically highlights:
GPT-6 Sol
in an internal factuality evaluation based on real conversations where users reported errors,
makes about
half as many errors as GPT-5.6 Sol.
But they also caution:
these were deliberately error-prone conversations,
not the overall typical error rate in general use.
So it’s inaccurate to say:
"GPT-6 Sol hallucinations dropped 50%."
More precisely:
errors roughly halved in this specific OpenAI error-inducing test.
Luna’s High Effort Performance Nearly Matches Old Sol
OpenAI also says:
GPT-6 Luna,
at higher effort levels,
comes close to GPT-5.6 Sol
in factuality tests,
but at roughly 1% of the cost.
If true,
it means many tasks previously requiring
mid-to-high tier models for reliable handling
could start shifting to
much cheaper model tiers.
This is the real
commercialization of model compression.
Not just making models smaller,
but lowering the price at equivalent capability levels.
Sol and Luna Don't Replace Astra
This is critical.
OpenAI’s official stance is clear:
GPT-6 Astra remains the most powerful overall model.
The most important projects,
most difficult research,
most complex computer use,
and most critical tasks
still belong to Astra.
The new strategy is more like:
large scale and simple:
Luna.
complex daily:
Sol.
extremely difficult and low volume:
Astra.
This aligns almost perfectly with what we wrote this afternoon about Claude’s business cases:
model routing.
Model Management Starts Resembling Company Workforce Allocation
You wouldn’t have your most senior engineer
spend every day organizing Excel files,
searching for documents,
or replying to simple emails.
Similarly,
having Astra doesn’t mean
every task requires Astra.
Enterprises may automatically route AI tasks by:
tier:
Luna,
routine processing:
Sol,
escalated to Astra
when difficult,
then downscale again after completion.
An AI model portfolio
starts to look like
workforce tiers.
What Will Happen to General ChatGPT Users?
At release,
OpenAI lets:
GPT-6 Sol
and
GPT-6 Luna
enter ChatGPT Work,
Codex,
and API,
with paid plans gaining phased access.
Free and Go users can use
GPT-6 Luna via desktop app.
The general ChatGPT interface adoption is gradual.
So, if you don’t see it yet,
it doesn’t mean it’s not released,
just that the rollout hasn’t reached your account.
Another Practical Change: Higher Usage Limits
With cheaper model serving,
OpenAI not only lowers API prices,
but the same compute budget
lets users run more.
They emphasize that Sol and Luna
bring
higher usage limits
and
more room to iterate.
This may feel more tangible than even a 3% increase in benchmark scores,
because for daily AI users,
the biggest annoyance isn’t
a model missing a math problem,
but
running into usage limits mid-task.
Why This "50% Price Cut" Is More Significant Than It Seems
If you use a model by
chatting ten times a day,
dropping input token cost from $4 to $2
barely registers.
But if a company
runs it continuously in an agent workflow,
it’s a different story.
For example, a monthly API bill of
$200,000
for the same task volume could theoretically drop significantly
just due to list price cuts.
Or the company might not save the money,
but instead let AI do twice the amount of work.
This is the real impact of price wars.
But We Can’t Declare "AI Costs Are Halved" Yet
Because halving token prices
doesn’t necessarily mean
your final task cost
halves.
Models might:
think longer,
call more tools,
produce more output,
have different workflows,
or because new models are better,
companies might assign
more complex tasks it wouldn’t have previously.
Final bills
may not decrease.
Thus, the true metric remains:
cost per completed task.
OpenAI and Anthropic Share This Unique Focus
Anthropic:
Opus 5.5.
OpenAI:
GPT-6 Sol/Luna.
Both releases spend substantial time telling you:
don’t just look at model price,
look at how much a real task costs,
how many iterations it takes,
failure rates,
whether you need to redo,
and how much context can be cached.
This means frontier AI
is moving into a new maturity stage.
Previously:
capability first.
Now:
capability-to-cost ratio.
This Price War Will Force Other Model Vendors to Follow
If high-end capabilities
can increasingly be shifted down to
mid-range and
even low-cost models,
pressure will spread to:
Google,
Meta,
Alibaba,
DeepSeek,
Mistral,
and other providers.
Because enterprise customers ultimately ask:
not
"What is your worldwide benchmark rank?"
but
"I have 5 million tasks a month—what will it cost to complete them?"
This question is far harsher than leaderboard rankings.
Open-weight Models Will Also Feel Pressure
One big appeal of open-weight models is:
cost.
You can deploy yourself,
avoiding always paying Frontier APIs’ high prices.
But if Frontier model APIs
keep getting cheaper,
enterprises will reevaluate:
buying GPUs,
maintaining infrastructure,
updating models,
monitoring security
versus
paying API fees directly.
So model price cuts
could even shift
the economic balance between
local AI,
cloud AI,
and open-weight AI.
But Cheapest Won't Always Win
Cheapness is
only one dimension.
Enterprises also consider:
reliability,
latency,
privacy,
data residency,
security,
tool ecosystems,
vendor lock-in,
availability,
model stability,
and most importantly,
how much human review is required.
If a model is 70% cheaper but employees spend
20 more minutes reviewing each time,
it might not be cost-effective.
AI procurement becomes
a real operations question.
Looking Back at This Morning, Today Was Special
This morning:
Anthropic told the market:
Claude Opus 5.5
can significantly reduce task costs.
Tonight:
OpenAI slashed prices further with
GPT-6 Sol and Luna.
In one day,
two leading frontier AI firms did not stop at
"We're smarter."
but began focusing on:
"We can complete work more cheaply."
This may be one of the most important turning points in the AI industry.
As intelligence itself becomes increasingly abundant,
what becomes truly scarce is:
how much reliable work a dollar can buy.
So the real significance of GPT-6 Sol and Luna
isn’t just OpenAI adding two model names,
but frontier AI moving
from capability races
to
efficiency races.
The true winner may not be
the one topping benchmarks,
but the one
delivering the most real work reliably at the lowest total cost.
Today, let's progress with AI a little.
Learn one AI skill daily.
Save a bit of time every day.
Improve a little every day.
SasaDaily, growing with you.
