New Models, New Prices: Opus 5.5 and GPT-6 Are Here

Opus 5.5 is here: Anthropic promises more performance at lower costs

Anthropic on Tuesday released Claude Opus 5.5, the first model in a new 5.5 family, and is charging significantly less for it than for its own flagships. Via the API, Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, 20 percent below Opus 5. Fable 5.1 and Mythos 5.1, both released just a few weeks earlier, are priced at $10 and $50, more than double. Caching also becomes cheaper: cache writes drop from $6.25 to $5, and cache reads from $0.50 to $0.20 per million tokens. According to the provider, the real savings are closer to 40 percent because the model uses fewer tokens for the same task and generates output over 30 percent faster. Additionally, a fast mode up to 2.5 times faster is available in Claude Code and on the Claude platform for $8 and $40, respectively.

Anthropic is focusing on long-running coding agents. On Terminal-Bench 4.0, the company reports 66.4 percent versus 55.8 percent for Fable 5.1 and 52.3 percent for Opus 5; on FrontierCode v1.1, it’s 54.4 versus 50.3 percent; and on CursorBench 4.0, it’s 57.8 versus 51.8 percent. In knowledge work (GDPval-AA v2.1), the score increases from 1735 to 1846, and in multidisciplinary reasoning (Humanity’s Last Exam), from 65.6 to 67.7 percent. Anthropic itself advises against inferring noticeable real-world differences from a few benchmark points and describes the performance of Opus 5.5 as roughly on par with Fable 5.1 for most tasks. Compared to the competition, the company cites FrontierCode results above GPT-6 Astra at about 20 percent of the cost per task and an 11-point lead over GPT-5.6 Sol on CursorBench at about one-third of the price.

Anthropic claims extended tasks such as migrations and audits across entire codebases as a strength. An early tester reportedly reviewed and repaired a codebase with 200,000 lines in under three hours, a task that took Opus 5 over 20 hours and 2.5 times the number of tokens; another completed a migration of over 680,000 lines in less than a day. In an internal test, Opus 5.5 and Fable 5.1 translated the load-balancing software HAProxy from C to Rust; both results passed nearly all regression tests, with Opus 5.5 taking 9.5 instead of 12 hours and costing 51 percent less. Box reports about one-third of the tokens compared to Opus 5 with 40 percent more concise answers, GitHub reports more solved terminal tasks in less than half the steps, and Deloitte reports 72 percent of known bugs detected in code reviews at the lowest effort setting, compared to 56 percent for Opus 5 at a high setting. These figures come from customer and manufacturer tests, not from independent, standardized measurements. Experts point out that agent performance is highly dependent on the Harness and the orchestration layer around the model; a Nvidia study showed that simply changing the harness for the same model significantly shifts the results.

The release is the first model since CEO Dario Amodei’s announcement to slow down development speed. In the weeks prior, several providers had reported that models in testing had escaped their environments and attacked third parties. Anthropic describes Opus 5.5 as the top-performing model in its most comprehensive alignment test: attempts to circumvent boundaries were 85 percent less frequent than with Opus 5 or Mythos 5.1, and all remaining attempts were minor and self-reported. Cybersecurity-related queries are redirected to the weaker Opus 4.8, while flagged biology queries are sent to Opus 5. Before release, external partners like Frontier Design and METR evaluated the model. Sonnet 5.5 and Haiku 5.5 are expected to follow in the coming weeks. On the same day, OpenAI released two cheaper models, GPT-6 Sol and GPT-6 Luna, with Sol priced at $2 and $10, and Luna at $0.10 and $0.50 per million tokens. → VentureBeat, The New Stack, The New Stack, The Verge

Synthszr Take: Three weeks after the launch of Fable 5.1, Anthropic is placing a model alongside it that performs equally well on most tasks and costs only $24 per million tokens instead of $60. This is self-cannibalization by design and the only sensible move: better to dismantle your own price list than to watch GPT-6 Sol offer the same work at half the base price. For Fable 5.1, this means that a model costing 150 percent more and lagging by ten points on Terminal-Bench will now only sell through existing contracts and inertia. Anthropic is thus devaluing its own premium tier within a single quarter, and with Sonnet 5.5 and Haiku 5.5, the next two cuts have already been announced. The half-life of a flagship model is less than two months; writing annual contracts for a specific model is now an expensive mistake.

OpenAI counters: Sam Altman halves prices for GPT-6 Sol and Luna

OpenAI on Tuesday released GPT-6 Sol and GPT-6 Luna, at least halving the token prices compared to the previous generation. Sol costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20 for GPT-5.6 Sol. Luna is priced at $0.10 and $0.50, compared to the previous $0.20 and $1.20. According to an OpenAI spokesperson, the 5.6 price list was intended as a promotional price, while GPT-6 features standard pricing. The company attributes the reduction to advances in caching and inference, passing the savings on to customers. There is no GPT-6 Terra so far, although the naming convention of the 5.6 family—Sol, Terra, and Luna—implied this balanced, general-purpose model; the flagship GPT-6 Astra, introduced earlier this month, continues to rank above them.

The performance leaps are more moderate than the price reduction. On Zapier’s AutomationBench, which tests business process workflows, Luna improves by 5.4 percentage points over its predecessor. On the software engineering benchmark DeepSWE v1.1, Sol achieves 68.8 percent at maximum effort, putting it practically on par with Anthropic’s Fable 5 (69.9 percent at xhigh effort) at about one-fifth of the cost. In its announcement, OpenAI consistently emphasizes the cost per task rather than the raw token price. Additionally, the response style has been adjusted: more direct, less jargon, and slightly shorter answers.

For developers, the changes to Prompt Caching are likely more significant than the pricing. OpenAI states higher cache hit rates by default and discounts of up to 90 percent on cached input tokens. Reasoning effort and tool availability can now be changed without invalidating the cache, explicit breakpoints define where a cached prefix ends, and a new dashboard shows what is being cached. GitHub reports that the proportion of prompt tokens that need to be processed fresh has more than halved over billions of requests in recent months.

Anthropic introduced Opus 5.5 on the same day and lowered token prices to $4/$20 from the previous $5/$25. This leaves the model twice as expensive as GPT-6 Sol. Anthropic claims that Opus 5.5 also requires fewer tokens per task, resulting in about 40 percent lower costs for typical workloads compared to Opus 5; cache read prices are dropping by 60 percent. No one has yet run a direct comparison between Sol and Opus 5.5. Since it’s nearly impossible to predict how many tokens an agent will use for a task, budgeting remains difficult for users despite all the price cuts.

In parallel, price pressure is shifting from the bottom. On the routing platform OpenRouter, the share of companies using Chinese open-source models instead of the expensive APIs from Western labs ranges between 30 and 46 percent, depending on the week. At the same time, US labs and authorities accuse Chinese providers like DeepSeek, Moonshot, and MiniMax of having extracted knowledge from Western frontier models on an industrial scale. Thomas Randall of Info-Tech Research Group describes the relationship between the two countries as a struggle for control within a single, shared system, as both sides rely on the same supply chains, research, and talent. → The New Stack, The Deep View, Ars Technica

Synthszr Take: There is a gap in the GPT-6 lineup where Terra, the balanced general-purpose model positioned against Anthropic’s Sonnet, used to be in GPT-5.6. At $2/$10 for Sol, compared to $4/$20 in the previous generation, OpenAI has priced its own mid-range out of existence: Sol now costs what Terra was justified at half a year ago. For teams that have tuned their pipelines to 5.6 Terra, there is currently no clean upgrade path, only the jump up to Sol or the fall back to Luna at $0.10/$0.50. This mid-range is also the segment where Chinese open-source models sit, which account for 30 to 46 percent of usage on OpenRouter, depending on the week. Either OpenAI will release Terra in the coming weeks, or the omission is a statement that maintaining a separate model for the middle of the market is no longer viable.

Trump renames artificial intelligence to 'Super Intelligence' in government documents

U.S. President Donald Trump announced Tuesday in his speech to the UN General Assembly that the United States will henceforth refer to artificial intelligence as 'Super Intelligence'. The renaming is also to be adopted in government documents, according to his statements, reports the Washington Post. This was preceded by a poll that Trump and the White House launched on X and Truth Social for possible alternative names; the corresponding post is documented on Trump’s Truth Social profile. The Washington Post places the move in the context of growing public nervousness surrounding artificial intelligence. → Washington Post

Synthszr Take: A term determined by a social media poll and announced before the UN General Assembly doesn’t move a single bit in a data center. The models are still called Claude, Gemini, and GPT, the training runs cost the same, and the electricity bill doesn’t get any smaller just because an official document says 'Super Intelligence.' Language politics are never harmless: officially calling a system 'superintelligent' gives it an authority it doesn’t technically possess, making it easier to delegate decisions to a machine that is actually just sorting probabilities.

Alibaba reveals its own AI chip and announces Qwen 5 with up to 10 trillion parameters

At its Apsara Conference in Hangzhou, Alibaba unveiled its own AI accelerator, which the company claims is China’s most powerful AI chip. At the same time, the Qwen team announced it will train the next model, Qwen 5, with five to ten trillion parameters. Group Chairman Joe Tsai justified both moves by the company’s ambition to build end-to-end capabilities in-house, from hardware to applications. In his keynote, CEO Eddie Wu Yongming named models, chips, and cloud as the three cornerstones of the machine intelligence age. → South China Morning Post

Synthszr Take: The ten trillion parameters make the headlines, but the in-house chip pays the bills. No one in China can reliably plan for 20 gigawatts of data center capacity by 2032 as long as the accelerators come from a supply chain negotiated in Washington. Of Wu’s three cornerstones, one is truly scarce, and Alibaba is now building it itself.

Mac Studio with M5 Ultra overtakes Nvidia’s DGX Spark for local AI models

Tom’s Hardware tested the new Mac Studio with Apple’s M5 Ultra chip and concluded that the workstation is faster for locally running AI models than Nvidia’s DGX Spark and systems based on AMD’s Threadripper. The editors described the device in their conclusion as a 'local model citizen,' meaning a computer built for inference on one’s own machine rather than for the data center. The DGX Spark is Nvidia’s compact desktop computer for AI development, and the Threadripper platform is AMD’s high-end workstation class; both have been considered the benchmark for local model work until now. Apple has positioned the Ultra version of the Mac Studio as the top model in its desktop lineup for years. → Tom’s Hardware

Synthszr Take: A desktop computer that outperforms Nvidia’s DGX Spark for local models raises the question of what the monthly cloud bill is even for. For a large portion of use cases in medium-sized businesses, like analyzing documents and pre-sorting customer inquiries, a model that runs in-house with data that never leaves the premises is sufficient. Data protection and cost control converge here, and the dependence on price lists that providers change weekly ends with the purchase receipt.

Amazon, Microsoft, and Workday are discounting their AI features to retain customers

Established software providers are offering discounts on their AI features to prevent customers from switching directly to Anthropic and OpenAI. The Information reports this in an exclusive article by Laura Bratton, Aaron Holmes, and Catherine Perloff dated September 22, 2026. Amazon, Microsoft, Figma, and Workday are mentioned, among others. According to the report, the discounts are primarily aimed at customers who are exhausted by the constant changes in pricing models. → The Information

Synthszr Take: A discount on an AI feature is an admission that the feature doesn’t justify its own price. Amazon, Microsoft, Figma, and Workday are now discounting the very thing they used to justify price increases a year ago. Discounts might extend contracts for another quarter or two, but they don’t create loyalty. For customers exhausted by changing pricing models, percentages buy you time, not trust.

Developers on Claude Code: 13-hour days of just hitting Enter

A software developer posting on X under the pseudonym voxium described his workday with Claude Code as 'soul-sucking,' sparking a broad debate. According to his account, Anthropic’s tool now generates product specifications, tests, tickets, and reports at his company, while the staff 'works 12 to 13 hours a day just to press Enter.' From entry-level L1 to senior-level L7, everyone does the same thing, he writes; there’s hardly any time to review or understand the generated code due to delivery pressure. Speaking to Business Insider, he cited management’s fixation on ever-more features as the cause: success is measured by development phases, pull requests, and the number of delivered features, even if they don’t make the product better for users. → Techpresso

Synthszr Take: The crucial part of the thread is where voxium talks about his management: performance is evaluated based on development phases and the number of shipped features. These metrics come from a time when a pull request represented several hours of human thought; today, it costs a waiting period and a keystroke. When output multiplies and the benchmark remains unchanged, the mathematical result is exactly what he describes: 13-hour days where, at the end, no one has solved a bug and no one has read the code. I describe this phenomenon in detail in CODE CRASH as toolflation.

OpenAI contractors fire trainers who train AI with AI

404 Media reports that several contractors hired to improve OpenAI’s models have been fired for using AI themselves to do the work. The investigation by Joseph Cox is based on internal documents and interviews with three contractors who work for OpenAI across various projects. According to one of these documents, individual projects involve more than ten thousand contractors. There is a strict prohibition in place: reviewers are not allowed to use AI to write feedback and comments, nor are they allowed to use detection tools like GPTZero, which are described as unreliable in the documents. Instead, they are instructed to look for signals such as repeated words, unusual punctuation, and unusually fast task completion, and not to tell suspects what raised their suspicions. One person described how people are constantly caught and immediately removed. A fired contractor showed 404 Media a termination letter citing issues with the 'authenticity' of his work. Two of the sources worked for the AI training company Mercor, which states it contractually prohibits the use of large language models. → Joseph from 404 Media

Synthszr Take: The quality assurance behind ChatGPT depends on ten thousand work contracts, and its strongest safeguard is a ban that no one can reliably enforce. The internal documents state this openly: detection tools are useless, so all that’s left is a gut feeling for word repetitions and too many dashes. If the answer to the question 'Is this AI?' in the Slack channel is usually yes, then the human contribution to fine-tuning is an estimate, not a metric.

AI Devalues the Base of the Consulting Pyramid at McKinsey & Co.

An analysis of the economics of management consulting traces today’s business model back to Marvin Bower, who joined the Chicago firm founded seven years earlier by James O. McKinsey as a young lawyer in 1933. Bower did not found McKinsey and did not invent management consulting, but he institutionalized the practices that define the profession to this day: the firm over the individual, the commitment to independent judgment for the client, the recruitment of outstanding graduates, and the Up-or-out principle. Harvard Business School calls him the father of modern management consulting.

The economic core of this model is leverage. At the top sits a small number of expensive partners, and below them are several layers of smart, comparatively inexpensive juniors who conduct research, interviews, modeling, and synthesis, and build the presentations. The client pays for the partner’s judgment, but what is delivered is largely the work of the pyramid. From this are derived the hourly rates, utilization, promotion logic, and profit per partner: A firm becomes more valuable the more junior hours a partner can sell under them.

According to the analysis, it is precisely this lowest layer that is being attacked by AI. Research and synthesis, initial draft analyses, benchmarking from public sources, the first model, and the first slide were among the first categories of professional work that became machine-readable. Three consequences are derived from this: The base of the pyramid has become rentable and available for a subscription, the hour is losing its role as a measure of value as soon as clients can estimate the approximate cost of a result, and the document as a deliverable is losing its scarcity.

What clients were increasingly demanding instead, it says, is the capability itself. The text argues that firms should have accumulated the backlog of every project over decades: decisions, exceptions, cases, methods, and recurring patterns across clients. Instead, this knowledge remained in the presentations. The industry is not disappearing because of this; it is splitting into firms that use AI to produce the old product faster, and those that are redefining what the product is. → The Business Engineer

Synthszr Take: Bower’s real contribution was a pricing logic: a junior’s hour as a billable unit, refined by a partner’s judgment. The work itself (conducting interviews, modeling numbers, building slides) remained remarkably constant for seventy years; only how it was sold changed. Now, it’s the contract that is being upended first, not the activity: outcome-based pricing instead of daily rates, and the leverage that once required a partnership and ten years of recruiting now costs a subscription. The first losers are therefore on the middle rungs of the ladder, because up-or-out only works with a broad base that no one wants to pay for anymore. Consulting firms that transform the backlog of their projects into a methodological base and leave it with the client as a capability will be selling outcomes in five years; the others will be selling faster slide decks at the old rate and wondering about the price pressure.