# GPT-6 vs Claude: prices and benchmarks across all four tiers

> Eight models, four tiers: compare intelligence scores, API prices and measured task costs for the latest OpenAI and Claude releases.

By BIG CHANGE Editorial

Published: 2026-09-22T19:53:39.563Z
Updated: 2026-09-22T22:39:13.693Z
Canonical: https://bigchange.ai/blog/gpt-6-luna-opus-5-5-price-release-race

![Orange token discs flow from a hopper toward a central switch, with a separate charcoal-and-ivory cartridge aligned to an orange-edged adapter.](https://bigchange.ai/api/media/file/gpt6-luna-opus-routing-hero-v1.png)
AI-generated conceptual illustration by BIG CHANGE. Cheap token volume and a premium model route are separate choices; a new model can also bring a compatibility connection to check. This conceptual illustration does not compare model quality or performance.

OpenAI and Anthropic released new models on September 22, but their lineups are moving at different speeds. GPT-6 Sol and Luna join Astra, while Terra is still on GPT-5.6. Claude Opus advances to 5.5, alongside Fable 5.1, Sonnet 5 and the nearly year-old Haiku 4.5.

The comparisons that matter are Astra against Fable, Sol against Opus, Terra against Sonnet, and Luna against Haiku. Looking across all four reveals a more interesting contest than either launch alone: Opus is challenging the flagship tier, Luna puts pressure on Anthropic's inexpensive offering, and the middle of both ranges remains unfinished.

## The big change

- **What changed:** Buyers can now get newer models at sharply different prices within the same provider's range. OpenAI has refreshed its inexpensive tier while Anthropic has concentrated this release on Opus, leaving Haiku almost a year behind in the release calendar.
- **Why it matters:** Product teams have more scope to cut the cost of routine AI use while paying for stronger results where mistakes are expensive. The difficult part is choosing by demonstrated performance: a model's place in the product range does not give a reliable ordering of its benchmark results.
- **What to watch:** Anthropic says Sonnet 5.5 and Haiku 5.5 will arrive in the coming weeks. Their prices and results will determine how much of today's opening for OpenAI survives, especially in products that make large numbers of inexpensive calls.

## The prices, tier by tier

All figures below are standard direct-API US dollars per million tokens, checked September 22. Cached input means a cache hit. OpenAI's listed rates apply through 272,000 input tokens; above that, full-request input rates double and output rates rise 50%. Cache writes, batch processing and tools have separate billing conditions. Sources: [OpenAI](https://developers.openai.com/api/docs/pricing) and [Anthropic](https://platform.claude.com/docs/en/about-claude/pricing).

| Model | Input | Cached input | Output |
| --- | --- | --- | --- |
| GPT-6 Astra | $10 | $1 | $50 |
| Claude Fable 5.1 | $10 | $0.25 | $50 |
| GPT-6 Sol | $2 | $0.20 | $10 |
| Claude Opus 5.5 | $4 | $0.20 | $20 |
| GPT-5.6 Terra | $2 | $0.20 | $12 |
| Claude Sonnet 5 | $2 | $0.20 | $10 |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 |
| Claude Haiku 4.5 | $1 | $0.10 | $5 |

### Intelligence and measured test spending

Artificial Analysis Intelligence Index v4.3.2, checked September 22. Costs are the operator’s weighted average spending per index task.

| Model | Tested setting | Intelligence index | Cost per index task |
| --- | --- | --- | --- |
| GPT-6 Astra | Max | 52.7 | $3.26 |
| Claude Fable 5.1 | Max with fallback | 53.4 | $7.63 |
| GPT-6 Sol | Max | 47.5 | Not listed |
| Claude Opus 5.5 | Max with fallback | 57.6 | $5.98 |
| GPT-5.6 Terra | Max | 42.1 | $1.40 |
| Claude Sonnet 5 | Max | 38.2 | $5.09 |
| GPT-6 Luna | Max | 37.3 | Not listed |
| Claude Haiku 4.5 | Reasoning enabled | 16.9 | $0.21 |

Source: [Artificial Analysis](https://artificialanalysis.ai/models); individual model results are linked in the comparisons below.

### Astra versus Fable: the flagship pair

Their standard input and output prices match. Fable's cache-hit rate is a quarter of Astra's, a difference that matters when applications repeatedly reuse eligible prompt content.

On Artificial Analysis's current Intelligence Index, Astra at max scores **52.7** and Fable 5.1 at max with fallback scores **53.4**. That is a close result. Artificial Analysis reports weighted spending of **$3.26 per index task for Astra**, against **$7.63 for Fable**. Astra delivers a close score at about 57% lower measured spending, despite their matching input/output rates. [Astra results](https://artificialanalysis.ai/models/gpt-6-astra), [Fable results](https://artificialanalysis.ai/models/claude-fable-5-1).

### Sol versus Opus: lower prices meet stronger scores

Sol's ordinary input and output rates are half Opus's; their cache-hit rates match. OpenAI also halved Sol's input and output prices from GPT-5.6 Sol's launch rates. Anthropic cut Opus's corresponding rates by 20% from Opus 5. [OpenAI launch](https://openai.com/index/introducing-gpt-6-sol-and-luna/), [Opus launch](https://www.anthropic.com/claude-opus-5-5).

Artificial Analysis gives Sol at max **47.5**, against **57.6** for Opus 5.5 at max with fallback. Opus leads Sol by 10.1 points while charging twice as much for ordinary input and output. It also scores above Astra and Fable at 60% lower input/output rates. Actual test spending changes that picture: Opus averages **$5.98 per index task**, about 84% more than Astra. Artificial Analysis's published cost chart did not yet provide Sol's figure when checked. [Sol results](https://artificialanalysis.ai/models/gpt-6-sol), [Opus results](https://artificialanalysis.ai/models/claude-opus-5-5).

### Terra versus Sonnet: the missing GPT-6 model

There is no GPT-6 Terra in OpenAI's current launch or model catalog. The available comparison is **GPT-5.6 Terra**, whose current [API documentation](https://developers.openai.com/api/docs/models/gpt-5.6-terra) confirms the rates above. OpenAI's [Work and Codex guide](https://learn.chatgpt.com/docs/models) also says the older 5.6 models remain available during the rollout.

At max effort, Terra scores **42.1** on Artificial Analysis's index and Sonnet 5 scores **38.2**. Terra also averages **$1.40 per index task**, versus **$5.09 for Sonnet**: a higher score at about 73% lower measured spending. Its output-token rate is 20% higher, showing how token consumption and caching can outweigh the sticker price. [Terra results](https://artificialanalysis.ai/models/gpt-5-6-terra), [Sonnet results](https://artificialanalysis.ai/models/claude-sonnet-5).

Sonnet 5 arrived on June 30. Anthropic has now announced Sonnet 5.5 for the coming weeks. A promised successor leaves today's Sonnet 5 comparison intact until the new version is available to test and price.

### Luna versus Haiku: the widest gap in release age

Luna costs one-tenth as much in each listed token category. Haiku 4.5 launched on October 15, 2025: **342 days before Luna**. It remains Anthropic's current released Haiku, with 5.5 promised in the coming weeks.

The performance gap also appears in independent testing. Artificial Analysis gives Luna at max **37.3**, compared with **16.9** for Haiku 4.5 with reasoning enabled. These are index points on the same current suite, alongside a 90% lower token price for Luna. Haiku reasoning averages **$0.21 per index task**; Luna’s corresponding cost was not yet listed in the operator’s published chart. [Luna results](https://artificialanalysis.ai/models/gpt-6-luna), [Haiku reasoning results](https://artificialanalysis.ai/models/claude-4-5-haiku-reasoning).

For Anthropic, this is an exposed part of the range: its cheapest current model faces a newly released competitor with both lower token rates and a much stronger result on that assessment. The announced Haiku update is therefore one of the next consequential releases for developers buying at volume.

The Artificial Analysis figures above use **Intelligence Index v4.3.2**, a ten-evaluation aggregate, checked September 22. Settings are named alongside the scores; Fable and Opus entries include the operator's fallback configuration. [Index and methodology](https://artificialanalysis.ai/models).

## What happens when agents have to finish work?

Zapier's **AutomationBench v1.0.6** tests end-to-end work in simulated business applications. A pass requires every scored outcome to be correct. Its September 22 leaderboard reports:

| Model | Tested setting | Pass rate | Cost per task |
| --- | --- | --- | --- |
| GPT-6 Astra | Max | 41.4% | $1.73 |
| Claude Fable 5.1 | Max, standalone | 22.37% | $2.45 |
| GPT-6 Sol | Xhigh | 33.2% | $0.27 |
| GPT-6 Sol | Max | 32.0% | $0.34 |
| Claude Opus 5.5 | Max | 40.0% | $1.28 |
| GPT-5.6 Terra | Max | 23.59% | $0.51 |
| Claude Sonnet 5 | Max | 10.65% | $0.88 |
| GPT-6 Luna | Max | 20.7% | $0.04 |
| Claude Haiku 4.5 | Not specified | 0.46% | $0.07 |

The separate Fable-plus-Opus-5 configuration scores 31.4%. Opus handles roughly 40% of its tasks; the displayed $2.45 excludes those fallback tokens.

Sol at xhigh spends about **79% less per attempt** than Opus at max for a 6.8-percentage-point lower pass rate. Luna beats Haiku at **43% lower task cost**. Astra leads this group; every pass rate remains below half. [Zapier's results and scoring method](https://zapier.com/benchmarks).

## The release calendar is busy and uneven

The companies' API release records provide two selected series. The OpenAI list follows numbered GPT versions through GPT-6 Astra. It excludes specialized Instant and Codex branches, and later tier expansions such as GPT-5.4 mini and nano on March 17 and GPT-6 Sol and Luna on September 22. The Anthropic list follows public Opus API versions. The “gap” is calendar days since the previous entry in that company’s selected series.

- **GPT-5 family — Aug. 7, 2025:** family launch; gap not applicable for the first entry.
- **GPT-5.1 — Nov. 13, 2025:** version update; 98-day gap.
- **GPT-5.2 — Dec. 11, 2025:** version update; 28-day gap.
- **GPT-5.4 — Mar. 5, 2026:** version update; 84-day gap.
- **GPT-5.5 — Apr. 24, 2026:** version update; 50-day gap.
- **GPT-5.6 — July 9, 2026:** family update with tiers; 76-day gap.
- **GPT-6 Astra — Sept. 3, 2026:** new family launch; 56-day gap.

Source: [OpenAI API changelog](https://developers.openai.com/api/docs/changelog). Separately, GPT-6 Sol and Luna arrived on September 22, 19 days after Astra. That interval measures a same-family tier expansion and is outside the numbered-version series above.

- **Opus 4 — May 22, 2025:** major version; gap not applicable for the first entry.
- **Opus 4.1 — Aug. 5, 2025:** point update; 75-day gap.
- **Opus 4.5 — Nov. 24, 2025:** version update; 111-day gap.
- **Opus 4.6 — Feb. 5, 2026:** version update; 73-day gap.
- **Opus 4.7 — Apr. 16, 2026:** version update; 70-day gap.
- **Opus 4.8 — May 28, 2026:** version update; 42-day gap.
- **Opus 5 — July 24, 2026:** major version; 57-day gap.
- **Opus 5.5 — Sept. 22, 2026:** version update and price cut; 60-day gap.

Source: [Anthropic API release notes](https://platform.claude.com/docs/en/release-notes/overview). Fable and restricted Mythos releases are separate branches, so they are not inserted into the Opus sequence.

OpenAI’s named-version gaps swing from 28 to 98 days; Anthropic’s Opus gaps range from 42 to 111 days. If “major release” means a change in the family number, GPT-5 to GPT-6 took 392 days and Opus 4 to Opus 5 took 428 days. The shorter gaps belong to point updates and tier expansions. Both series fluctuate: shorter gaps are followed by longer ones.

## The practical pressure is to upgrade selectively

The launch calendar keeps creating reasons to reconsider a model choice. Today's four pairs show why a single provider-wide verdict would miss useful differences: Anthropic has a strong new Opus, OpenAI has a newly competitive inexpensive tier, and both companies still have older models occupying parts of their ranges.

Changing versions also takes engineering work. Anthropic's [Opus 5.5 migration guide](https://platform.claude.com/docs/en/models/opus-5-5/whats-new-opus-5-5) says thinking is always on, forced tool selection returns an error, and the earlier computer-use tool type is rejected on the Claude API and Google Cloud. Existing integrations using those options need changes.

The next announced development is Anthropic's Sonnet and Haiku refresh. That gives buyers a concrete reason to revisit the lower tiers soon. For now, the price lists and independent results already show where the competition has shifted, and where an older model still deserves consideration.

## Sources

- [OpenAI API pricing](https://developers.openai.com/api/docs/pricing) — Standard input, cache-hit and output rates and long-context bands used throughout the four-tier comparison. API rates are separate from subscriptions.
- [OpenAI API changelog](https://developers.openai.com/api/docs/changelog) — Dated API entries support the OpenAI release series. GPT-6 Astra is treated as the family launch; Sol and Luna on September 22 are classified separately as a tier expansion rather than another generation.
- [OpenAI GPT-6 Luna model documentation](https://developers.openai.com/api/docs/models/gpt-6-luna) — Luna’s documented intended workload, context window and long-prompt pricing conditions used in the comparison.
- [Anthropic API pricing](https://platform.claude.com/docs/en/about-claude/pricing) — Current Fable, Opus, Sonnet and Haiku API prices, including cache-hit rates; supports the within-line Opus price reduction.
- [Anthropic API release notes](https://platform.claude.com/docs/en/release-notes/overview) — Dated launch entries support the Opus version sequence. Fable and restricted Mythos releases are excluded because they are separate branches, not generally available Opus successors.
- [Anthropic: What's new in Claude Opus 5.5](https://platform.claude.com/docs/en/models/opus-5-5/whats-new-opus-5-5) — The migration documentation identifies breaking API changes, including always-on thinking, errors for forced tool choice, and rejection of the earlier computer-use tool type on the Claude API and Google Cloud. These are documented compatibility costs, not a hypothetical integration story.
- [Big Change: Claude Opus 5.5 makes long AI jobs cheaper, with limits on what it will handle](https://bigchange.ai/blog/claude-opus-5-5-cost-performance-access) — Our separate Opus launch analysis supplies further detail on access and cost. This comparison adds four-tier prices, independent results and release chronology.
- [Anthropic: Claude Haiku 4.5](https://platform.claude.com/docs/en/models/haiku-4-5/overview) — Haiku’s speed-oriented positioning, pricing and 200,000-token context window support the comparison’s scope and its limits.
- [OpenAI: Work and Codex models](https://learn.chatgpt.com/docs/models) — Lists GPT-5.6 Terra among models remaining available in Work and Codex during the GPT-6 Sol and Luna rollout, subject to account access.
- [OpenAI API model catalog](https://developers.openai.com/api/docs/models) — Provides OpenAI’s intended roles for Astra, Sol and Luna and their place in the current API lineup.
- [OpenAI: Introducing GPT-6 Sol and Luna](https://openai.com/index/introducing-gpt-6-sol-and-luna/) — September 22 launch and Sol price change. Its launch charts include earlier Claude versions, so current cross-provider scores here come from benchmark operators.
- [Anthropic: Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5) — September 22 Opus launch, price cut and announcement that Sonnet 5.5 and Haiku 5.5 are coming in the following weeks.
- [OpenAI: GPT-5.6 Terra](https://developers.openai.com/api/docs/models/gpt-5.6-terra) — Current Terra API rates corroborate the available 5.6 model; no GPT-6 Terra is introduced by this source.
- [Artificial Analysis: Intelligence Index](https://artificialanalysis.ai/models) — Independent v4.3.2 index checked September 22. Model pages supply scores, settings and weighted cost per index task; Sol and Luna cost entries were unavailable. Fable and Opus include fallback.
- [Zapier: AutomationBench](https://zapier.com/benchmarks) — Version 1.0.6 results, task costs and strict scoring. Fable standalone and fallback results are distinguished; fallback costs are excluded from the displayed combination price.
- [Anthropic: Claude Haiku](https://www.anthropic.com/claude/haiku) — Official Haiku 4.5 launch date of October 15, 2025, used to calculate its age at the September 22 comparison.