Opus 5.5 Crushes Fable & Astra

English
CChase AI
Computing/SoftwareBusiness NewsInternet Technology

Transcript

00:00:00So we got a brand new Opus model.
00:00:02We've gone all the way from Claude 5 and jumped to 5.5.
00:00:06But what sort of changes do you need to know about?
00:00:08Well, the highlight is that Anthropic is saying Claude Opus 5.5
00:00:12essentially operates at the same level as Fable 5.1, but costs 40% less.
00:00:20So we're basically getting a cheaper Fable 5.1.
00:00:24Now to start, let's take a look at the benchmarks.
00:00:26We have Opus 5.5 compared to Fable 5.1, Opus 5, as well as GPT-6 Astra and 5.6 Sol.
00:00:33Across the board, Opus 5.5 is showing numbers that beat pretty much everybody.
00:00:38The only two places it loses are to GPT-6 Astra,
00:00:42and that's on business workflows and agentic scientific research.
00:00:45On the business workflows benchmark, it loses by 1.4%.
00:00:50And on the scientific research benchmark, it loses by about 6%.
00:00:54But when we compare Opus 5.5 to Fable 5.1, and again, it is 40% cheaper,
00:01:00it beats it pretty much everywhere.
00:01:01And on top of that, we see huge leaps when we compare it to the Opus 5 model
00:01:04and pretty much everything related to coding.
00:01:07But when we look at benchmarks, it doesn't really do us much good
00:01:09just to see the performance levels.
00:01:11We need to see what it looks like at different effort levels
00:01:14and what it looks like in terms of cost per task.
00:01:16Because in theory, we can say, okay, per million tokens, Opus costs this,
00:01:21but that doesn't matter if it uses way more tokens to complete a task.
00:01:24So what we have right here is the Terminal Bench 4.0,
00:01:26and we have Opus 5.5 in the red.
00:01:29We have Fable 5.1 in the green, and Astra over here in the gray.
00:01:32Now, purely in terms of performance, we see that Opus 5.5 is beating pretty much everybody,
00:01:37especially when we look at high, extra high, and max.
00:01:40And in terms of price, it also looks like it's fairly efficient.
00:01:44When we look at it at high, we're scoring 64.2% at $3.88.
00:01:50For Astra on high, it's $7.21.
00:01:53So, you know, almost twice as expensive.
00:01:56And then when we compare it to Fable, 5.1 on high, not only does it perform way worse,
00:02:0149.4% versus 64.2, it's also way more expensive.
00:02:06At $10.50, again, compared to $3.88.
00:02:10So, in general, Opus 5.5 looks like it performs better and is more efficient.
00:02:17And when we look at the Frontier Code and Cursor Bench benchmarks,
00:02:20it kind of plays out the same way.
00:02:22It's performing better at virtually every single effort level,
00:02:25and at each effort level, it also does it cheaper.
00:02:28Now, when we look at these benchmarks, what we do kind of see,
00:02:31and we've noticed this across a whole lot of different models,
00:02:34is that we tend to get diminishing returns as we go from high all the way to extra high and max.
00:02:39When we compare highs, output 64.2 versus 64.8, pretty much the same,
00:02:45yet the cost is almost four times as much.
00:02:49We see the same thing with Cursor Bench.
00:02:51Again, high gives us pretty much max outputs at a way cheaper cost.
00:02:54And it's even more exaggerated on the Frontier Code set,
00:02:57where medium effort level actually gives us the best performance at 54.6 compared to max at 54.4,
00:03:05yet it is six, seven times cheaper, which is wild.
00:03:09I mean, this one's really crazy because Opus 5.5 on medium outperforms every single model out there
00:03:14while being way cheaper than every other model's low setting, which is crazy.
00:03:18And when we take a look at the knowledge work benchmarks, again, we see pretty much the same thing play out,
00:03:23the one exception being the automation bench, where Astra does pull ahead.
00:03:27Now, in terms of its straight up token cost per million, Opus 5.5 is $4 per input and $20 per output,
00:03:35which is cheaper than Opus 5.
00:03:37On top of that, we have cheaper cash reads and writes, which is huge.
00:03:40For reference, Astra and Fable 5.1, their input is $10 and their output is $50.
00:03:46So if these benchmarks are to be taken at face value, then we have an extremely efficient model on our hands.
00:03:51Now, one of the big issues with Opus 5 was how it communicated.
00:03:55We pretty much had the Opus slot mean running around, where it was extremely jargon heavy,
00:04:00and it was difficult to even understand what it was saying.
00:04:03This is something Anthropic has heard, and they're telling us that they have solved this problem with Opus 5.5,
00:04:09making it a much better collaborator, because we didn't actually understand what it's saying.
00:04:13Specifically, they say it is less likely to use jargon or idiosyncratic phrases,
00:04:17and it follows the writing rules you give it.
00:04:20And they have some comparisons here, where it shows what Opus 5 would tell us,
00:04:23versus what Opus 5.5 would tell us, whether that's a bug, or summarizing a thread,
00:04:27or explaining a design change.
00:04:29And again, if you've used Opus 5 at all, you understand.
00:04:32It is very annoying to actually talk to it.
00:04:34And lastly, we have safeguards.
00:04:36Anthropic is telling us that Opus 5.5 pretty much falls in line
00:04:38with the Fable series in terms of like cybersecurity and biology type questions.
00:04:43So if it deems that you're asking questions about cybersecurity that you're going to use
00:04:47in a negative fashion, like you're trying to use it to like ask somebody,
00:04:52it's going to reroute your question.
00:04:53And when it reroutes it, it's going to reroute it to Opus 4.8.
00:04:57This also applies to biology.
00:04:59So in general, if you can't ask Fable 5.1 this question,
00:05:02it reroutes you with Fable.
00:05:03It's going to reroute you when you ask the same question to Opus 5.5.
00:05:07Also similar to Fable is how Opus 5.5 handles distillation.
00:05:11It stopped API users from editing Claude's prior context in an attempt to extract Claude's reasoning.
00:05:16So same exact sort of method to stop these attacks.
00:05:21And lastly, in terms of availability, it is available everywhere today.
00:05:25So on paper, what's not to love here?
00:05:27We have a model that is just as good as Fable and is cheaper.
00:05:31Now, will it hold up in reality?
00:05:33Well, it's just going to take time.
00:05:34We're going to have to play around to see if that's the case.
00:05:36But if you've been on Twitter at all lately,
00:05:39you have seen some of the outputs people have been getting with Opus
00:05:42when it's been silently routed to Opus 5.5 over the last week or so.
00:05:46And the results have been impressive.
00:05:48So I'm excited for this model because honestly,
00:05:51it is a pain in the butt to use the Anthropic plans,
00:05:54especially when they have reduced our usage of late.
00:05:56And we can only use half of our weekly allowance on Fable.
00:05:59And so your second half almost has felt wasted with Opus 5.
00:06:02But if I can pretty much get Fable 5.1 and then Opus 5.5,
00:06:06which is basically Fable for the other 50%,
00:06:08well, then I'm not going to complain.
00:06:10And that's big because for a lot of people, it's been like,
00:06:12well, why don't we just go over to OpenAI and chat GPT and start using Astra?
00:06:15Because they don't limit us in that sense.
00:06:17Well, if you're able to get a $200 plan, that is.
00:06:20So really excited for this one.
00:06:22Definitely check it out and let me know what you think.

Key Takeaway

Claude Opus 5.5 delivers performance that matches or exceeds Fable 5.1 and GPT-6 Astra across major coding and knowledge benchmarks while cutting token costs by 40% to 60%.

Highlights

  • Claude Opus 5.5 performs at the same level as Fable 5.1 while costing 40% less.

  • Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, compared to $10 input and $50 output for GPT-6 Astra and Fable 5.1.

  • On Terminal Bench 4.0 at high effort, Opus 5.5 scored 64.2% at $3.88 per task, beating Fable 5.1's 49.4% score at $10.50.

  • Opus 5.5 on medium effort achieves a 54.6% score on Frontier Code, outperforming every model's max setting while being cheaper than low settings.

  • Requests regarding cybersecurity or biology deemed malicious are automatically rerouted to Opus 4.8.

Timeline

Opus 5.5 Overview and Core Benchmarks

  • Claude Opus 5.5 operates at Fable 5.1 performance levels at a 40% discount.
  • Opus 5.5 outperforms Fable 5.1, Opus 5, and GPT-5.6 Sol across nearly all standard benchmarks.
  • GPT-6 Astra retains a narrow lead over Opus 5.5 in business workflows by 1.4% and agentic scientific research by 6%.

Anthropic updated the Opus line directly from version 5 to 5.5. The new model matches top-tier capabilities while drastically dropping operating expenses. Across general benchmark suites, it shows significant leaps over Opus 5, particularly in software development tasks.

Cost Efficiency and Effort Level Performance

  • Terminal Bench 4.0 testing shows Opus 5.5 achieving 64.2% accuracy at high effort for $3.88, compared to Astra's $7.21 and Fable 5.1's $10.50.
  • Higher effort settings yield diminishing returns, with high effort scoring 64.2% and max effort scoring 64.8% at nearly four times the price.
  • Medium effort on Frontier Code delivers a top score of 54.6% for six to seven times lower cost than max effort.

Evaluating performance per task demonstrates that higher compute budgets do not scale linearly with accuracy. On Frontier Code and Cursor Bench, medium and high effort settings produce optimal outputs without the exponential cost surge of max effort settings. The medium effort tier on Opus 5.5 routinely beats competitor models running at maximum compute.

Pricing Structure and Communication Improvements

  • Opus 5.5 rates are $4 per million input tokens and $20 per million output tokens.
  • Anthropic reduced jargon and idiosyncratic phrasing to make output easier to read.
  • The model follows user-defined style and writing rules more precisely than Opus 5.

Direct token costs place Opus 5.5 substantially below Astra and Fable 5.1, which both charge $10 input and $50 output per million tokens. In addition to lower base rates, prompt caching reads and writes are cheaper. Natural language output fixes address common complaints about excessive technical jargon present in previous iterations.

Safety Rules, Safeguards, and Platform Availability

  • Potentially harmful cybersecurity or biological queries are automatically rerouted to Opus 4.8.
  • API modifications that alter historical context to extract internal reasoning are blocked.
  • Opus 5.5 is live across all platforms immediately.

Safety boundaries align with the protocols built into Fable 5.1, restricting misuse in sensitive domain areas through immediate query redirection. Distillation prevention mechanisms prevent users from manipulating prior conversational history to siphon chain-of-thought data. The model is fully deployed, providing a viable alternative for subscribers facing weekly usage caps on Fable.

Community Posts

No posts yet. Be the first to write about this video!

Write about this video