The 2026 Rate Slide: An ItInsightsWorld Record of Every LLM Price Move by OpenAI and Anthropic

Frontier text-model rates fell so repeatedly during 2026 that no single press release explains the shift; only the sequence does. This ItInsightsWorld record lays out each dated announcement as one connected market process rather than a string of isolated promotions. Every figure that follows is quoted in the order input MTok / output MTok / cache read MTok.

Announcements in date order

June 26, 2026 — a three-tier family arrives

OpenAI introduced GPT-5.6 Sol and Luna. Sol inherited GPT 5.5's slot at an unchanged $5 / $30 / $0.50, while Luna covered cheaper use cases as the small model at $1 / $6 / $0.10. Terra was announced in the gap between the two and attracted little attention.

June 30, 2026 — a launch discount labelled temporary

Anthropic announced Claude Sonnet 5 and positioned it as closer to its Opus models, though the reception was mixed. The launch rate of $2 / $10 / $0.20 was presented initially as temporary and slated to rise.

July 24, 2026 — Opus returns to flagship parity

Anthropic unveiled Claude Opus 5, a major update after Opus 4.5's effect on agentic coding, at $5 / $25 / $0.50 — close to GPT-5.6 Sol. Some benchmarks surpassed the much larger Fable 5, but the real-world response remained mixed.

July 30, 2026 — the first deep cut

OpenAI lowered GPT-5.6 Luna by 80% to $0.20 / $1.20 / $0.02. Read as model quality against cost-per-task, the arithmetic implied that without the cut the model would have trailed DeepSeek V4 Pro.

August 10, 2026 — temporary becomes permanent

Anthropic made Sonnet 5's discounted launch price permanent, reversing its own stated plan to raise it later.

August 21, 2026 — the flagship moves

OpenAI said it would cut GPT-5.6 Sol by ~20%, to $4 / $20 / $0.40, for three months.

September 1, 2026 — a new flagship with a cache discount attached

Anthropic announced Fable 5.1 at $10 / $50 / $0.25. It beat Fable 5 from June in all benchmark areas and cut cache read from $10 / $50 / $1, a reduction of 75% on that line alone.

Cache read is where agent budgets actually go

For coding agents, cache read accounts for 96% - 99% of token spend, because the cache carries the whole prior conversation each turn. Cache is hit around 90 - 100% of the time depending on inference infrastructure, so a 75% cache cut implies 70-75% lower total cost.

September 3, 2026 — matched performance, unmatched cache line

OpenAI announced GPT-6 Astra, similar in size to Fable 5.1 and priced at $10 / $50 / $1, but lacking the cheaper cache read. Astra's performance is comparable to Fable.

September 22, 2026 — both labs cut on one day

• OpenAI announced GPT-6 Sol at $2 / $10 / $0.20, halving input and output relative to the temporarily discounted GPT-5.6 Sol. GPT-6 Luna also came at half the prior price: $0.10 / $0.50 / $0.01.
• The same day Anthropic announced Claude Opus 5.5, saying explicitly that “Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that”. Its $4 / $20 / $0.20 total sits far below Opus 5 because of a 20% input/output cut plus a bigger cache cut, lowering total cost 40-60% while beating Fable 5.1.

September 29, 2026 — a flagship revised within a week

GPT-6 Sol proved underwhelming in performance, so a week later OpenAI released GPT-6.1 Sol. It matched Astra's performance, as personal use confirmed, and OpenAI cut the cache price in half to $2 / $10 / $0.10, matching Anthropic's 5%-of-input cache read cost and making the family cumulatively 75% cheaper.

The same sequence as a table

Date	Model	Input / Output / Cache read	Move recorded
June 26, 2026	GPT-5.6 Sol	$5 / $30 / $0.50	unchanged from GPT 5.5
June 26, 2026	Luna	$1 / $6 / $0.10	small-model tier introduced
June 30, 2026	Claude Sonnet 5	$2 / $10 / $0.20	launch price, later made permanent
July 24, 2026	Claude Opus 5	$5 / $25 / $0.50	parity with the OpenAI flagship
July 30, 2026	GPT-5.6 Luna	$0.20 / $1.20 / $0.02	80% reduction
August 21, 2026	GPT-5.6 Sol	$4 / $20 / $0.40	~20% cut for three months
September 1, 2026	Fable 5.1	$10 / $50 / $0.25	cache read cut 75% from $10 / $50 / $1
September 3, 2026	GPT-6 Astra	$10 / $50 / $1	no equivalent cache discount
September 22, 2026	GPT-6 Sol	$2 / $10 / $0.20	input and output halved
September 22, 2026	GPT-6 Luna	$0.10 / $0.50 / $0.01	half the prior price
September 22, 2026	Claude Opus 5.5	$4 / $20 / $0.20	total cost 40-60% below Opus 5
September 29, 2026	GPT-6.1 Sol	$2 / $10 / $0.10	cache halved; family cumulatively 75% cheaper
April 2026	DeepSeek V4 Flash 0420	$0.14 / $0.28 / $0.0028	2% cache read proportion
October 24th, 2026	DeepSeek Flash V4.1 1017	$0.15 / $0.60 / $0.003	29 distinct providers listed
just a few days ago	Kimi K3.1 Pro	$2 / $10 / $0.20	below Kimi K3 at $3 / $15 / $0.30

What the aggregate numbers show

Measured from the GPT-5.6 launches in June to the present, the cost of using popular LLMs has cratered — a slide that even I, watching this space a year ago, did not forecast. Subscriptions for agentic harnesses such as Codex and Claude Code benefit as well, and the models do not appear to have compromised output quality to reach these prices. Data from Vercel's AI model gateway supports decreasing cost-per-token down to 43.3% of the April 2026 cost, while aggregate spend rose because models improve faster than prices fall.

Behavioural economics: why these models price like commodities

If you test agents against each new LLM, subjective vibes are often the only differentiator. Since last November triggered hypergrowth of agentic LLMs, models behave more like commodities: switching is trivial through a harness like OpenCode or Zed Agent, or via a different API endpoint string through OpenRouter, leaving little moat. There are many competitors beyond Google and Meta, offering sufficient substitutes and creating pricing pressure.

Open weights set the floor

Consider DeepSeek Flash V4.1, including its updated 1017 variant; the listing on OpenRouter was retrieved October 24th, 2026. The original DeepSeek V4 Flash 0420 appeared in April 2026 at $0.14 / $0.28 / $0.0028 and went megaviral partly because OpenAI lacked a simple-task small model. The key figure was $0.0028, because a cache read proportion of 2% is far below the 5%-10% of OpenAI and Anthropic models, lowering effective cost beyond the listed price. I suspect the July cut of 80% on GPT-5.6 Luna targeted these smaller open-weight models. Note the operative words “from DeepSeek”. On scoring, DeepSeek Flash V4.1 1017 now performs significantly better than DeepSeek Flash V4 — 45 on the Artificial Analysis index versus 24 — while DeepSeek still charges $0.15 / $0.60 / $0.003.

On OpenRouter, any inference provider can list a hosted model; requests through the OpenRouter API are routed to available providers by default, or by preferences such as “lowest price”. Back in May, there were 13 distinct providers hosting DeepSeek V4 Flash. For the updated 1017 variant, the sorted list of providers for DeepSeek V4.1 Flash 1017 shows 29 distinct providers, retrieved October 24th, 2026.

When effective input costs are considered, with cache read and hit rates interacting, many providers are close. Cloudflare and DeepInfra were within 10% of DeepSeek's cost, though these cheapest models still hold only a minority of total token share.

The differentiation spend the closed labs cannot stop

OpenAI and Anthropic need to differentiate from open-weight alternatives with higher-quality models; they are doing so, especially with Opus 5.5, but must keep doing that, which is why billions of dollars are going into R&D. The open-weight field is catching up: just a few days ago, Moonshot AI announced Kimi K3.1 Pro, comparable to Opus 5.5 — Kimi at 55 versus Opus 5.5 at 58 — with half the input/output price and the same cache read: $2 / $10 / $0.20. It shares Kimi K3's size and architecture, and Kimi K3 is hosted by over 20 inference providers. Once weights drop on November 4th, most can be expected to support the new model at comparable prices. Moonshot AI also joined the price cuts, pricing Kimi K3.1 Pro below the original Kimi K3 at $3 / $15 / $0.30.

Durable low prices rather than later hikes

AI detractors often say tokens are priced low only to hook users and will later rise to “the true cost of a token”, wrecking boosters. The analogy is Uber and Lyft in the 2010s, but unlike that duopoly era, OpenAI and Anthropic cannot solely set prices and inference entry barriers are lower. If many inference providers offer similar quality at similar prices and switching costs are negligible, prices may keep falling more. That is why both companies seek moats, such as enterprise commitments and personal agentic assistants like OpenAI Dots and Claude Duet.

The notion of a sudden token-price jump conflicts with behavioural economics and the observed significant declines for better models, not only from OpenAI and Anthropic. As long as scalers charge below revenue per token, providers will reduce costs to compete until OpenAI and Anthropic must follow. The 29 distinct providers on OpenRouter have little reason to serve an unprofitable model; they would stop and redirect capacity, making “the true cost of a token” possibly equal to marginal cost.

IPO pumping

Both OpenAI and Anthropic are close to IPOing, which raises questions about the price cuts. OpenAI is rumored to IPO in 2027, and Anthropic within a couple months. Leaked financials show they are losing billions from R&D and scaling, though the reports do not capture recent agentic expansion. Investors will examine negative profits and demand plans.

Why cut a primary revenue source by up to 80%? That seems opposite to looking attractive in an S-1 prospectus. Yet inference profit margins may differ; some reports say profit-per-token margins are as healthy as 80% with caveats. How much margin remains after the cuts?

Personally, the price cuts changed my agent spending differently from what the labs may expect. I used to hold a $100/month subscription to Codex, the 5x Pro plan, when GPT-5.6 Sol was the top model, and I exhausted all included usage despite annoying random resets. When GPT-6.1 Sol launched and effective usage cost dropped by 75%, I cut the max usage quota to 1/5th, because the same model was 4x more efficient, roughly neutral.

Subscription margins for Claude Code and Codex remain uncertain because their value may exceed API usage; if capital tightens, I suspect subscription allowances would be cut before base prices. At the same GPT-6.1 Sol announcement, OpenAI reduced the usage allowance by 50% for the $200/month tier in Codex under the argument “to reflect our increasingly efficient models” — technically correct, though funny to put in writing. I will be curious to see Anthropic's IPO prospectus and public inference margins soon.

Model obsolescence and infrastructure

It may not benefit OpenAI or Anthropic to make pricier models obsolete intentionally. After Claude Opus 5.5 released a month ago, there has been little discussion of Fable 5.1 or GPT-6 Astra, because Opus 5.5 is just that good, and a top-and-left view from Anthropic reinforced that position. I expect Fable 5.2 or GPT-6.1 Astra to release in a couple weeks and prove me wrong, and I will be curious how each is priced.

A less cynical explanation is that LLM infrastructure has improved. When Anthropic said “Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that”, I believe it. If efficiency reduces the need for massive data centres, that could cause cascading economic failures, but likely not, because inference is only one part of the compute mix and data centres remain needed for the pipeline.

Reading the record as a mechanism

Set out in sequence, the year stops resembling a run of promotions and starts resembling an automated response: one dated announcement, a same-day or one-week counter-move, and a price floor redrawn by whoever is willing to host weights. The cuts remain odd, and I am certainly not complaining. Open-weight models and secondary inference providers will continue to act as an economic check on LLM pricing and on other business manoeuvres the large labs may concoct; even if the Anthropic and OpenAI listings end badly enough to fold both companies, the LLM and agent industry as a whole would not die with them.