Is GPT-6.1 Sol cheaper than Claude Opus 5.5?

Yes. Per token it is half the price ($2/$10 vs $4/$20 per million). Artificial Analysis measures Opus 5.5 at 4.2x to 8.9x the cost per task at the same effort level.

Did Anthropic cut Claude Code weekly limits?

Both readings are true. From Sep 14 the limit is 25% above the pre-promotion level, permanently, but 17% below the 50%-boosted level used May 13 to Sep 13.

Was Claude Opus 5.5 nerfed?

Unproven. There is no official confirmation and no quality incident on the status page from Sep 1 to Oct 7. One independent test found no drop but can only detect about 6.6 points.

Which effort level should I use?

Both models default to medium. The top setting costs 5x to 11x the lowest for 10 to 16 index points, so compare models at equal effort and move up only where the extra points pay for themselves.

GPT-6.1 Sol vs Opus 5.5: Complaints Fact-Checked

Updated October 13, 2026 · first published October 13, 2026

Quick answer: Two sets of complaints are running at once. On the OpenAI side: Pro $200 got cut, GPT-6.1 Sol feels worse, Astra eats the allowance. On the Anthropic side: Claude Code limits shrank, Opus 5.5 burns...

Two sets of complaints are running at once. On the OpenAI side: Pro $200 got cut, GPT-6.1 Sol feels worse, Astra eats the allowance. On the Anthropic side: Claude Code limits shrank, Opus 5.5 burns the weekly cap, Opus 5.5 got nerfed. We took each claim and checked it against the vendors' own pages, Artificial Analysis, the GitHub issues people filed with numbers, and our own logs. Most of the money complaints hold up. The "it got stupid" complaints do not have evidence yet.

The short version:

Is GPT cheaper than Claude? Yes, and it is the best-supported claim in the whole debate. Per token GPT-6.1 Sol is half the price of Opus 5.5 ($2/$10 vs $4/$20). Per task, Artificial Analysis measures Opus 5.5 at 4.2x to 8.9x the cost, depending on effort level.
Did Claude cut its limits? Both stories are true. From September 14 the weekly Claude Code limit is 25% above the pre-promotion level, permanently. It is also 17% below the 50%-boosted level people had all summer (150 to 125).
Did OpenAI cut Pro $200? Yes, from 20x to 10x Plus on October 30, price unchanged. We covered it in the full repricing; here we only check it against the rest.
Is Opus 5.5 nerfed? Is GPT-6.1 Sol stupid? Unproven either way. The one controlled test running so far has found no drop, but it admits it could miss a small one. No incident on Anthropic's status page matches a quality problem.
What is the real cost driver? Effort level. On both models the top setting costs 5x to 11x the lowest, for 10 to 16 index points.

Which complaints hold up?

Complaint	Verdict	Evidence
OpenAI halved Pro $200	True	20x to 10x Plus on Oct 30, GPT-6 Pro chat 200 to 100 a week. Reported by several outlets and an OpenAI email quoted on Hacker News. We could not open OpenAI's help page (it returns 403 to our fetcher).
Claude limits were cut	True and false	Anthropic's support page: +25% over pre-promo, permanent. Anthropic's own wording: a 17% reduction from the boosted level.
Opus 5.5 burns the weekly cap	Mixed	One user measured 3.1x faster burn after the Sep 25 reset. Others report using 1% to 20% of a plan. Workloads differ and are unstated.
Opus 5.5 was nerfed	Unproven	No official confirmation, no status-page quality incident (Sep 1 to Oct 7), one independent test with no drop so far.
GPT-6.1 Sol is stupid	Unproven	Artificial Analysis scores it one point below Astra. One user calls it "a train wreck in my project". No shared task.
GPT is cheaper	True	List prices, Artificial Analysis cost per task, and our own logs all agree.

What did Anthropic change in the Claude Code limits?

The weekly Claude Code limit moved three times in five months. Anthropic ran a 50% boost from May 13 to September 13. From September 14 the limit is 25% above the level before the boost, and Anthropic says that is permanent. Indexing the pre-boost limit at 100, it went 100, then 150, then 125. That is a drop of 16.7% from where users stood in early September, which Anthropic itself called a 17% reduction when it announced the change on August 29. The five-hour limits were not touched.

Date	Claude Code weekly limit (pre-promo = 100)
Before May 13, 2026	100
May 13 to Sep 13, 2026	150 (temporary boost)
From Sep 14, 2026	125 (permanent)
Sep 22, 2026	Opus 5.5 launches; reported higher 5-hour limits and a one-off reset that expires Oct 22

Anthropic publishes no hours or token caps for any plan. Its pricing page says only that Max 5x and 20x are 5x and 20x Pro per five-hour session, with a weekly cap on top, and that Max has a separate weekly limit for Fable models. Any "hours per week" figure you see is third-party. We found no official October change to Claude limits or prices.

What does each model cost per token?

These are list prices per million tokens, read on each vendor's own pricing page on October 13.

Model	Input	Cached input	Output
GPT-6 Astra	$10	$1.00	$50
GPT-6.1 Sol	$2	$0.10	$10
GPT-6 Sol	$2	$0.20	$10
GPT-6 Luna	$0.10	$0.01	$0.50
Claude Fable 5.1	$10	$0.25	$50
Claude Opus 5	$5	$0.50	$25
Claude Opus 5.5	$4	$0.20	$20
Claude Sonnet 5.5	$2	$0.10	$10
Claude Haiku 5.5 (prompts to 100K)	$0.10	$0.01	$0.50

Three details change the bill more than the headline rate:

Long prompts cost more on OpenAI. Above 272K input tokens, GPT-6.1 Sol doubles to $4 in and rises to $15 out. Haiku 5.5 steps up fivefold above 100K.
Opus 5.5 is 20% cheaper per token than Opus 5. Anthropic claims about 40% lower cost on typical workloads. The newer Claude tokenizer produces about 30% more tokens for the same text, so a token price comparison across generations understates the real change.
Cache reads are where coding agents live. GPT-6.1 Sol halved its cached rate from $0.20 to $0.10. That is half of Opus 5.5's $0.20. Our Codex logs showed 98% of input tokens cached, which is why the cached rate matters more than the input rate.

How much does effort level change the cost?

Effort is the biggest lever and the least discussed one. Both models take five levels: low, medium, high, xhigh, max. Both default to medium. Artificial Analysis ran both across its Intelligence Index and published score and cost per task for every level.

Effort	Opus 5.5 score	Opus 5.5 $ / task	6.1 Sol score	6.1 Sol $ / task	Opus costs
low	42	$0.55	42	$0.13	4.2x
medium	51	$1.34	48	$0.21	6.4x
high	54	$1.82	50	$0.32	5.7x
xhigh	56	$3.46	51	$0.39	8.9x
max	58	$5.98	52	$0.72	8.3x

Three readings of the table:

Sol is cheaper at every effort level. At the same index score of 42, Sol costs $0.13 and Opus $0.55. Opus 5.5 pulls ahead only when you let it think: its max setting scores 6 points above Sol's max, at 8.3x the cost.
The last steps are the expensive ones. Opus goes from 42 to 58 for 10.9x the cost. Sol goes from 42 to 52 for 5.5x. On Sol, max burns about 7x the output tokens of low (67M against 9.0M across the index) for 10 points.
Price per token explains only part of the gap. Opus 5.5 is 2x Sol per token. A secondary write-up puts Opus at 260M output tokens against Sol's 67M for the index, about 3.9x. Multiply 2 by 3.9 and you get 7.8x, close to the measured 8.3x. Opus is dearer both because each token costs more and because it uses more of them. Opus 5.5 also cannot switch thinking off.

Our own coding measurements sit at the low end of that range: GPT-6 Sol needed 3.3x to 4.2x fewer dollars than Claude per task in our cost-per-task article. A general benchmark index and a coding workload will not match exactly, but the direction and the order of magnitude agree.

What do the Codex plans actually give?

OpenAI's pricing page lists local Codex messages per five-hour window on Plus, and the model you pick moves the number a lot:

Model on Plus	Messages per 5 hours
GPT-6 Luna	350 to 3,000
GPT-6.1 Sol	15 to 160
GPT-6 Sol	15 to 150
GPT-6 Astra	5 to 45

Astra gets three to four times fewer messages than Sol, against a 5x price gap. That fits the "Astra drained my window" complaints without any nerf: the allowance is simply denominated so that the expensive model costs more of it. One Hacker News user wrote that Astra uses about 7x the allowance of 6.1 Sol in his testing. That is one person with no stated workload, but it sits in the same range. The pricing page also states that Fast mode draws 2.5x included usage and Ultrafast 8x. Pro plans currently carry no five-hour limit, only a weekly one, and OpenAI publishes no token budget for any tier.

What did people measure themselves?

We looked for people who posted numbers instead of feelings. The useful ones are on GitHub, in the anthropics/claude-code issue tracker. We could not read Reddit or X, so those threads are not covered here.

Issue	Claim	Method and caveat
#97398 (Sep 26)	Weekly burn rose after the Sep 25 reset: 93 responses per 1% before, 30 after	Local transcripts, deduplicated. That is 3.1x on the user's own figures, though the title says 3.6x. One user, one workload.
#97074 (Sep 25)	Headless `claude -p` burns about 1.8x more of the five-hour window per token than interactive	About $1.6 per 1% headless against $3.0 interactive. Same workload.
#98679 (Oct 1)	Opus 5.5 median thinking tokens 67, then 126; median output 447, then 711	Small sample (8 sessions in the second period), and the CLI version changed at the same time.
#96241 (Sep 23)	Ultracode used 70% of a Max 5x plan in 20 minutes	18 agents running at once. One anecdote with a number.

The first and third point the same way: more tokens per request drains the weekly cap faster. They are different users and different mechanisms, so treat that as a loose match, not a replication. On the other side, Hacker News commenters report using at most 1% of a $20 weekly limit for a code port, or never passing 20% of a plan. A 1% reading and a 54%-of-Max-in-two-days reading are both probably true. The workloads are not the same.

Is anyone testing for a nerf?

One project is. Livenerf runs a frozen panel of 78 questions with 90 samples a day for 30 days. On October 11 it was on day 18 and had published only its baseline: no drop detected. Its own documentation says it can detect a change of about 6.6 points per 10 days, and that it could not tell Opus 5 from Opus 5.5 at 99% confidence. So "no drop" does not prove "no nerf". It rules out a large one.

Anthropic's history is the other relevant fact. Its April 23 postmortem named three real causes of Claude Code quality complaints: default effort cut from high to medium (Mar 4 to Apr 7), a bug that cleared thinking history every turn (Mar 26 to Apr 10), and a prompt that capped replies at 25 words between tool calls (Apr 16 to Apr 20). The thinking-history bug probably also drained limits faster, through cache misses. Anthropic reset limits for all subscribers. So "it feels worse" was right in the spring, and the cause was configuration, not a quieter model. That is why we would check the effort setting and the client version before blaming the weights. The Anthropic status page shows no quality incident between September 1 and October 7; the entries are errors, latency and a spend-limit bug.

For GPT-6.1 Sol the picture is thinner. OpenAI says it scores near Astra at one fifth of the token price, and Artificial Analysis puts it one index point below Astra. The complaints we found are single posts, and several early "nerf" posts were about GPT-6 Sol, before 6.1 shipped on September 29. See why silent downgrades break your cost model for what a real regression test needs.

How should you check your own plan?

Log the percentage, not the feeling. Codex writes used_percent and token counts into ~/.codex/sessions/. Price the requests at list rates, then divide dollars by percentage points used and multiply by 100. That is your dollars per window, the method behind our 60-day comparison.
Record effort and client version with every reading. Issue #98679 shows why: a version bump and a date change look identical in the data.
Keep interactive and headless runs separate. Issue #97074 measured 1.8x between them.
Compare at equal effort. Medium against max is a 5x to 8x swing on either model.
Re-measure after every reset or launch. Both labs moved limits around model launches this autumn.

What we could not verify?

OpenAI's help-center pages for the Pro change returned 403 to our fetcher; the Oct 30 numbers rest on multiple outlets and a quoted OpenAI email, not on a page we read.
Neither lab publishes weekly caps in tokens or dollars. Every "hours per week" number is third-party.
The 260M versus 67M output-token comparison and the Opus 5.5 launch-limit increase come from secondary write-ups.
We found no like-for-like SWE-bench, METR, Aider or LMArena figures for these two models, and no ccusage-based public tracker.
Vendor benchmark tables (DeepSWE, Terminal-Bench, OSWorld) are vendor-reported and do not always list the same competitor models.
Reddit and X were not searched, so the loudest complaint venues are missing.

Sources

Want this applied to your stack? Bring the provider bills, gateway logs, and top workflows; we will map the cost drivers and savings path. Book a free audit →

Back to finopsllm.com