OpenAI Raised the Price 150%. It Made One Job 50% Cheaper and Another 75% More Expensive.

The number every launch-day headline led with

GPT-6 Astra shipped on September 3, 2026, priced at $10 per million input tokens and $50 per million output tokens, with cached input at $1.00 per million. That is a flat 2.5x increase over GPT-5.6 Sol's $4/$20 rate, confirmed independently by both OpenAI's own published rate card and third-party launch coverage. Almost every writeup led with that number, and most framed it the way price increases from a market leader usually get framed: pay more, get more.

The same day, Artificial Analysis published its own scores for the model. On the Intelligence Index, its composite of 9 evaluations spanning math, science, coding, and reasoning, Astra scored 61. Out of 202 models tracked on the index, that placed it 8th. Not first. Eighth, tied exactly with the model it was built to replace.

What 61 actually ties

GPT-5.6 Sol, at its own max effort setting, also scores 61 on the Intelligence Index. Claude Fable 5.1, the model currently leading that board, scores 66, 5 points ahead. Meta's Muse Spark 1.3 also sits ahead of Astra on the same index. A 150% price increase bought OpenAI a tie with its own previous flagship, not a leap past the competition.

Token efficiency is the one place Astra clearly improved on general intelligence work: Artificial Analysis measured roughly 10% fewer output tokens than Sol needed to answer the same Intelligence Index questions. That is a real gain. It just is not close to enough to offset a 2.5x price increase. Multiply the two together the naive way, price ratio times a token-count ratio alone, and you would expect something in the neighborhood of a 2.25x cost increase. Artificial Analysis's own weighted cost-per-task methodology, which accounts for input tokens and the discounted cached-token tier alongside output tokens, lands the real number at 75% more expensive per completed task than Sol, not 125% more. The token savings blunt the price hike. They do not come close to erasing it.

# The naive read: price alone
sol_output_price = 20        # $ / 1M output tokens
astra_output_price = 50      # $ / 1M output tokens
price_ratio = astra_output_price / sol_output_price
print(f"Sticker price alone: Astra costs {price_ratio:.1f}x Sol per token")
# Sticker price alone: Astra costs 2.5x Sol per token

# Artificial Analysis's own weighted cost-per-completed-task figure
# (accounts for input + cached + output tokens, not output count alone):
cost_per_task_increase = 0.75   # AA's reported figure: "75% more expensive per task"
print(f"Real cost per completed Intelligence Index task: +{cost_per_task_increase:.0%}")
# Real cost per completed Intelligence Index task: +75%
# The ~10% output-token savings blunts the 2.5x sticker price. It doesn't erase it.

Now switch the job to coding, and the verdict flips

Everything above is general-purpose intelligence work: reasoning, science, knowledge tasks. Hand the identical model, at the identical new price, a coding task instead, and the story inverts. On Artificial Analysis's Coding Agent Index, Astra scores 67, approximately tying Claude Opus 5 and Claude Fable 5. Claude Fable 5.1 still leads outright at 70.

The efficiency gap that was a rounding error on intelligence work is enormous on coding work. To hit that score of 67, Astra needs roughly one-third the tokens GPT-5.6 Sol needs at its own max setting, and roughly one-fifth what Claude Opus 5 needs at its highest (xhigh) effort setting, 70% more token-efficient than Sol on this specific workload, in Artificial Analysis's own framing.

*GPT-5.6 Sol's Coding Agent Index score was not separately published; the 3x token figure is Artificial Analysis's own stated multiplier for Astra's efficiency gain over Sol on coding tasks specifically. Claude Fable 5's exact token count on this benchmark was not published; Artificial Analysis states only that Astra costs less than half of Fable 5's per-task price at the identical score.

Run the two numbers you can multiply directly, price ratio times token ratio, against Claude Opus 5, and the result matches what Artificial Analysis reports independently: Astra lands at well under half of Opus 5's cost for the same completed coding task.

# Astra vs. Claude Opus 5, coding tasks only
astra_price_per_m_out = 50     # $
opus5_price_per_m_out = 25     # $
price_ratio = astra_price_per_m_out / opus5_price_per_m_out      # 2.0 (Astra costs 2x per token)

astra_tokens_per_task = 1
opus5_tokens_per_task = 5      # Astra uses "one fifth" the tokens Opus 5 (xhigh) needs
token_ratio = astra_tokens_per_task / opus5_tokens_per_task       # 0.2

cost_ratio = price_ratio * token_ratio
print(f"Astra costs {cost_ratio:.0%} of Opus 5's price per completed coding task")
# Astra costs 40% of Opus 5's price per completed coding task
# -> 60% cheaper, comfortably "less than half the cost," matching Artificial Analysis's finding

Same invoice line. Same model. Same 2.5x price hike. Depending entirely on what you asked it to do, that price hike is either a 75% markup or a 50%-plus discount, per completed task.

The hallucination number that isn't what it looks like

One more figure from the release is worth slowing down on. Astra's hallucination rate on Artificial Analysis's AA-Omniscience evaluation falls from 92% to 51% at max effort, a drop that reads, on its own, like a dramatic safety and reliability win. Raw accuracy on that same benchmark climbs only 4 points. The arithmetic only closes one way: most of the hallucination-rate improvement comes from the model abstaining, saying it does not know, more often, not from it knowing more. A model that answers less and hedges more will always show a better hallucination rate on paper, whether or not it got smarter.

It is not a clean upgrade anywhere else, either

Astra gained 6 points on Humanity's Last Exam and roughly 80 Elo points on Artificial Analysis's AA-Briefcase evaluation. It also gave ground back on other parts of the same Intelligence Index: GDPval-AA v2, built from real business and knowledge-work tasks, drops by roughly 80 Elo points. Banking support (τ³-Banking), scientific coding (SciCode), and long-context reasoning (AA-LCR) each fall 2 to 3 points.

Astra did not get uniformly smarter for the price. It got differently shaped, stronger on some evaluations, measurably weaker on others, inside the same composite index that produced its one headline number.

The number that actually matters, and the one that doesn't ship in the press release

None of this makes GPT-6 Astra a bad model. It is a genuinely strong coding model at a genuinely competitive real-world price, once you account for how few tokens it burns to get there. It is also, on general intelligence work, a worse deal than the model it replaced, once you account for the same thing. Both facts are true, they come from the identical price change, and neither one is visible from the sticker price alone.

The launch-day framing, price up 150%, model up too, treated “price per token” as a stand-in for value. It isn't one. Cost per completed task is the only number that tells you what a model actually costs to use, and that number does not show up in a rate card or a press release. It has to be measured, per workload, every time a lab ships something calling itself the next generational leap.

If you are routing real traffic instead of reading headlines

The practical failure mode here is common and avoidable: hard-coding a routing decision around a launch-day price or a single leaderboard number, then never re-checking it once the workload mix changes. GPT-6 Astra is proof that the same price change can be a bargain and a ripoff simultaneously, depending only on what you send it. MegaBrain routes every call across 500+ models at zero markup and surfaces real, current cost-per-call data instead of a launch-day price frozen in a blog post, so when a model like this one is a steal for your coding agents and a markup for your general-knowledge endpoints, that shows up as a number you can route around, not a headline you have to remember to re-check every time a new model ships.

Sign up at getmegabrain.com to route across GPT-6 Astra and 500+ other models with real, per-call cost visibility, broken down by the workload you are actually running, not the one on the rate card.