2026 AI API Pricing Showdown: Who's Cheapest (Based on Official Pricing + Real Measurements)

TL;DR: As of May 2026, Claude Sonnet 4.6 ($3 input / $15 output per MTok) still wins on the performance/cost ratio. Output is 60% the price of Claude Opus 4.6, on par with GPT-5, and it scores 73.4% on SWE-bench Verified — only 4.6 points below Opus. For a typical coding turn (5K input / 2K output), Sonnet 4.6 costs ≈ $0.045 per turn (about ¥0.32 at the 7.2 ¥/USD official rate). Through CCSub's Starter tier (0.7× multiplier + ¥1 = $1 balance), the actual RMB cost drops to ≈ ¥0.03 per turn — roughly 9.4% of direct API pricing. (New here? See what a Claude API proxy is and how to set one up.)

Sources and methodology

Every official price below comes from the vendor's published pricing page as of May 2026:

• Anthropic Pricing: https://www.anthropic.com/pricing#api (last updated 2026-04)
• OpenAI API Pricing: https://openai.com/api/pricing/ (last updated 2026-04)
• Google AI (Gemini) Pricing: https://ai.google.dev/pricing (last updated 2026-03)
• AWS Bedrock Pricing: https://aws.amazon.com/bedrock/pricing/ (last updated 2026-04)

Benchmark numbers come from each vendor's published model cards plus independent third-party leaderboards:

• SWE-bench Verified (real-world code repair): https://www.swebench.com/
• GPQA Diamond (graduate-level QA): from the Anthropic Claude 4.6 model card and OpenAI GPT-5 system card
• Aider polyglot (multi-language code agent): https://aider.chat/docs/leaderboards/

Pricing changes over time; always cross-reference the official pages above before quoting these numbers.

Mainstream model pricing as of 2026-05 (USD per million tokens)

| Model | Input | Output | Cache Read | Context | Source | |---|---|---|---|---|---| | Claude Opus 4.6 | $5.00 | $25.00 | $0.50 (10%) | 1M | Anthropic | | Claude Sonnet 4.6 | $3.00 | $15.00 | $0.30 (10%) | 1M | Anthropic | | Claude Haiku 4.5 | $0.80 | $4.00 | $0.08 (10%) | 200K | Anthropic | | GPT-5 | $3.00 | $15.00 | n/a | 400K | OpenAI | | GPT-5 Codex | $5.00 | $20.00 | n/a | 400K | OpenAI | | GPT-5.4 | $5.00 | $15.00 | n/a | 400K | OpenAI | | GPT-4o | $2.50 | $10.00 | n/a | 128K | OpenAI | | o3 | $10.00 | $40.00 | n/a | 200K | OpenAI | | o4-mini | $1.10 | $4.40 | n/a | 200K | OpenAI | | Gemini 2.5 Pro | $1.25 | $10.00 | n/a | 2M | Google | | Bedrock Claude Opus 4.6 | $5.00 | $25.00 | $0.50 (10%) | 1M | AWS |

Performance benchmarks (higher is better)

| Model | SWE-bench Verified | GPQA Diamond | Aider Polyglot | |---|---|---|---| | Claude Opus 4.6 | 78.0% | 86.5% | 78.2% | | Claude Sonnet 4.6 | 73.4% | 81.5% | 71.4% | | GPT-5 (high reasoning) | 74.9% | 88.1% | 65.4% | | GPT-5 Codex | 74.5% | 79.0% | 70.0% | | Gemini 2.5 Pro | 67.2% | 84.0% | 59.0% | | Claude Haiku 4.5 | 51.0% | 65.0% | 35.5% |

Numbers come from vendor model cards and independent leaderboards as of 2026-05. Each benchmark fluctuates ±2pp depending on prompt and temperature; trust the official page for the latest values.

Measured: cost of a single coding turn

I ran 200 real coding tasks through the claude CLI (bug fixes, component scaffolding, refactors, repo reads) and tracked the per-turn token usage:

• Average input: 4,800 tokens (history + system prompt + tool calls)
• Average output: 2,100 tokens (thinking trace + code blocks + commentary)
• Cache hit rate: 62% (system prompt + project context reuse)

Calculating direct via Anthropic at Sonnet 4.6 with prompt caching enabled:

Non-cache input: 4,800 × 0.38 ≈ 1,800 tokens × $3.00  = $0.0054
Cache input:     4,800 × 0.62 ≈ 3,000 tokens × $0.30  = $0.0009
Output:          2,100 tokens × $15.00                = $0.0315
─────────────────────────────────────────────────────────────
Total per turn:  ≈ $0.0378  ≈ ¥0.27 at official 7.2 FX

Actual RMB cost via CCSub

CCSub stacks two layers of discount for Chinese users:

1. FX layer: ¥1 buys $1 of API balance (vs official ~7.2 ¥/USD)
2. Multiplier layer: Starter 0.7×, Pro 0.8×, Claude Max shared 1.5×, Business 2.5×, Bedrock 3.6×

Same Sonnet 4.6 workload routed through the cheapest Starter tier (0.7×):

Direct Anthropic (RMB):  $0.0378 × 7.2          = ¥0.272
CCSub Starter (RMB):     $0.0378 × 1 × 0.7      = ¥0.0265
─────────────────────────────────────────────────────────
Savings:                 (0.272 − 0.0265) / 0.272 = 90.3%

Over 1,000 daily coding turns: direct ≈ ¥272, CCSub Starter ≈ ¥27. The savings cover ~9,000 additional turns.

Best plan by user profile

| Profile | Monthly usage | Recommended plan | Estimated monthly cost | |---|---|---|---| | Occasional scripting / learning | < $10 | PAYGO ¥20 starter | ¥10–20 | | Daily development (< $15/day) | $15/day × 22 workdays | Monthly Entry ¥199 | ¥199 | | Heavy Claude Code user | $50/day × 30 | Monthly Standard ¥499 | ¥499 | | Full-time indie dev | $120/day × 30 | Monthly Advanced ¥1,188 | ¥1,188 | | Small team (2–5 people) | $200/day × 30 | Monthly Team ¥1,888 | ¥1,888 | | Enterprise prod | $500+/day, needs SLA | Bedrock Enterprise | Custom |

Monthly vs PAYGO decision tree

Is your daily USD usage steady?
├── No (< $5/day or highly variable)
│   └── Pick PAYGO — never expires
└── Yes, ≥ $10/day
    └── Pick a monthly plan — as low as ≈ 0.31 ¥/USD
        ├── < $20/day   → Entry ¥199
        ├── < $50/day   → Standard ¥499 (recommended)
        ├── < $120/day  → Advanced ¥1,188
        └── ≥ $200/day  → Team / Business

VIP cumulative discount (auto-upgrade)

CCSub auto-upgrades VIP based on cumulative top-up. The multiplier discount applies to every model:

| Tier | Multiplier | Cumulative top-up | Combined discount | |------|------|------------|---------| | Normal | 1.00× | Free | 13.9% (FX layer only) | | VIP1 | 0.98× | ¥299 | 13.6% | | VIP3 | 0.93× | ¥2,999 | 12.9% | | VIP8 | 0.88× | ¥9,999 | 12.2% |

A VIP8 Sonnet 4.6 turn costs ¥0.0235 — about 8.6% of the official price.

Why prompt caching matters more in China

Anthropic's cache_control mechanism prices repeated prompt sections at 10% of input cost. For Claude Code-style workloads with long system prompts (typically 5K–15K tokens), a warm cache cuts input cost another 70–80%.

In practice: across 100 consecutive turns in the same project, the cache hit rate peaked at 89%, lowering total cost by another 35%.

CCSub passes cache_control through untouched and bills cache reads at 10% of input price. Most domestic relays drop this header — if your relay strips it, your Claude Code bill ends up higher than going direct.

Conclusions

1. Best performance/cost: Claude Sonnet 4.6 (73.4% SWE-bench, $3/$15 per MTok) — the right default for Claude Code.
2. Reach for Opus when it matters: Claude Opus 4.6 (78.0% SWE-bench) for architecture decisions and long agent runs.
3. Cheap classifier / formatter: Claude Haiku 4.5 or GPT-4o — use for triage, formatting, lightweight agents.
4. Best path for users in China: CCSub Starter (0.7×) + Claude Sonnet 4.6 + prompt caching enabled → ~90% cheaper than direct Anthropic.
5. Stable heavy usage: Monthly Standard ¥499 + Sonnet 4.6 → $50/day balance covers a full-time engineer's daily Claude Code use.

---

References

1. Anthropic Pricing — https://www.anthropic.com/pricing#api (accessed 2026-05-06)
2. OpenAI API Pricing — https://openai.com/api/pricing/ (accessed 2026-05-06)
3. Google AI Pricing — https://ai.google.dev/pricing (accessed 2026-05-06)
4. AWS Bedrock Pricing — https://aws.amazon.com/bedrock/pricing/ (accessed 2026-05-06)
5. SWE-bench Verified Leaderboard — https://www.swebench.com/ (accessed 2026-05-06)
6. Aider Polyglot Leaderboard — https://aider.chat/docs/leaderboards/ (accessed 2026-05-06)
7. CCSub Pricing — https://www.ccsub.net/pricing

Start saving → CCSub Sign up