Between June and August 2026, Anthropic shipped Claude Fable 5, Sonnet 5 and Opus 5, OpenAI released GPT-5.6 and Google added Gemini 3.7 Flash. Nobody won outright: capable models got cheaper, but prices, names and versions now change every few weeks.
The summer 2026 releases did not produce a clear winner. What they changed for a business is the price and pace of capable AI. Anthropic shipped Claude Fable 5, Sonnet 5 and Opus 5 between 9 June and 24 July, OpenAI made GPT-5.6 generally available on 9 July, and Google released Gemini 3.7 Flash on 13 August. Mid-priced models moved closer to the flagships, list prices fell, and model names, versions and introductory prices now change every few weeks.
It was also not one August event. There is no single model called Claude 5, Google's headline August release was a Flash model rather than a new Pro, and GPT-5.6 reached customers in July. For a small or mid-sized business the dates matter less than three shifts: cheaper mid-tier models, agent features built in as standard, and faster turnover of the versions you build on.
- Claude Sonnet 5 per million input / output tokens, against $3 / $15 for Sonnet 4.6
- $2 / $10
- Claude Sonnet 5 per million input / output tokens, against $3 / $15 for Sonnet 4.6
- price cut on OpenAI's GPT-5.6 Luna from 30 July
- 80%
- price cut on OpenAI's GPT-5.6 Luna from 30 July
- between Gemini 3.6 Flash and its successor, Gemini 3.7 Flash
- 3 weeks
- between Gemini 3.6 Flash and its successor, Gemini 3.7 Flash
- more tokens, approximately, for the same text on newer Claude models
- 30%
- more tokens, approximately, for the same text on newer Claude models
01
What actually shipped, and when
Read the names carefully. Claude 5 is a family, not a model: Fable 5 launched as the most capable Claude open to all customers, with Opus 5 and Sonnet 5 below it on price. GPT-5.6 is three models under one number: Sol for frontier work, Terra for a balance of intelligence and cost, and Luna for high-volume jobs. On Google's developer API, the newest Pro model was still Gemini 3.1 Pro, in preview since February.
9 Jun
Claude Fable 5
Anthropic's most capable model for all customers, at $10 / $50 per million tokens, with a 1 million token context window.
26 Jun
GPT-5.6 limited preview
At the US government's request, OpenAI first gave access to a small group of trusted partners.
30 Jun
Claude Sonnet 5
Launched at an introductory $2 / $10, made the standard price on 10 August.
9 Jul
GPT-5.6 Sol, Terra and Luna
General availability in ChatGPT and the API.
24 Jul
Claude Opus 5
Billed as a step-change over Opus 4.8, at the same $5 / $25.
30 Jul
GPT-5.6 price cuts
Luna 80% cheaper, Terra 20% cheaper.
13 Aug
Gemini 3.7 Flash
Three weeks after Gemini 3.6 Flash, at an introductory $0.75 / $3.75 until 31 December.
21 Aug
GPT-5.6 Sol price cut
Down to $4 / $20, a promotional price OpenAI says will run at least until 21 November.
02
Prices fell, but read the small print
On paper, capable AI got cheaper again. These are list prices on 1 September 2026, per million tokens, the chunks of text a model reads and writes.
| Model | Released | Input / output per million tokens |
|---|---|---|
| Claude Fable 5 | 9 June | $10 / $50 |
| Claude Opus 5 | 24 July | $5 / $25 |
| Claude Sonnet 5 | 30 June | $2 / $10 |
| GPT-5.6 Sol | 9 July | $4 / $20, promotional |
| GPT-5.6 Terra | 9 July | $2 / $12 |
| GPT-5.6 Luna | 9 July | $0.20 / $1.20 |
| Gemini 3.7 Flash | 13 August | $0.75 / $3.75, introductory |
Three details change what you actually pay. First, tokenisers: Anthropic says Sonnet 5's new tokeniser can turn the same text into roughly 1.0 to 1.35 times as many tokens, so at the top of that range the saving over Sonnet 4.6 shrinks from a third to about a tenth. Second, promotions end: Gemini 3.7 Flash rises to $1.50 / $7.50 on 1 January 2027, and Sol's promotional price is guaranteed only until 21 November. Third, aliases: OpenAI's plain gpt-5.6 name routes requests to Sol, the most expensive tier.
03
Agent features are now standard equipment
Every release in the wave was pitched at multi-step work rather than single answers. Anthropic says Sonnet 5 can make plans and use tools such as browsers and terminals, and it became the default model on Claude's Free and Pro plans. OpenAI's GPT-5.6 added programmatic tool calling and, in beta, multi-agent orchestration to its API. Google says Gemini 3.7 Flash plans multi-step tasks and tool calls more diligently than 3.6 Flash, and it scored 34% against 22% for 3.6 Flash on GDP.pdf, a test of reading complex documents.
Images are now baseline input: Claude reads text and images, and GPT-5.6 accepts images at their original size. Audio is not universal, though. Claude's current models take text and images but not sound, so a voice workflow still needs a transcription step in front.
our most intelligent workhorse model yet for coding and agents
Safety rules tightened too. OpenAI rated all three GPT-5.6 models High for cybersecurity and for biological and chemical risk under its Preparedness Framework, the first time the smaller, faster models in one of its families got that label. Claude Fable 5 runs safety classifiers that can refuse a request, and it requires 30-day data retention, so it is not available to customers who need zero retention. Access can change at short notice, too: after the US government applied export controls to Fable 5 on 12 June, Anthropic suspended it for all users, restoring access on 1 July. If you handle confidential client data, check each model's terms, not just the provider's.
04
Churn is the new normal, so pin your versions
Google followed Gemini 3.6 Flash with 3.7 Flash three weeks later. Anthropic moved from Sonnet 4.6 in February to Sonnet 5 at the end of June. OpenAI's GPT-5.6 arrived about 11 weeks after GPT-5.5. A tool built on whatever the latest alias points to will change behaviour under you.
Pinning means naming the exact model in your code, such as claude-sonnet-5, rather than a moving alias. Anthropic treats every Claude model ID as a fixed snapshot and has committed not to retire Sonnet 5 before 30 June 2027. New versions can also break old code: on Sonnet 5, setting sampling controls such as temperature to non-default values returns an error, and the older manual thinking mode is gone. An integration that ran on Sonnet 4.6 can fail if someone simply swaps the model name.
05
What to do this quarter
You do not need to follow every launch. You need one working process and a habit of checking it. For the mid-tier option in detail, see our piece on Claude Sonnet 5 and agentic work; if data must stay on your own servers, read open-weight versus API models.
- Pick one workflow with clear inputs and a human checkpoint, such as enquiry triage, invoice reading or quote drafting, and measure hours saved per month.
- Start mid-tier. Try Sonnet 5, GPT-5.6 Terra or Gemini 3.7 Flash first, and pay for Fable 5 or Sol only where a cheaper model fails your test.
- Price it on your own data, using measured token counts and the price that applies after any promotion ends.
- Pin the model ID in your code or your vendor's statement of work, and note its retirement date.
- Check data terms for confidential work, because retention rules differ by model.
- Re-test every quarter on the same task set before you switch.
Sources
Facts current as of 1 September 2026. List prices are in US dollars and change often, so check each provider's pricing page before you budget.