개요
Claude Sonnet 4.8 represents an anticipated intermediate release in Anthropic's Sonnet lineage, positioned between the February 2026 Sonnet 4.6 and what ultimately became Claude Sonnet 5 (released June 30, 2026). Based on leaked internal documentation and code references examined in May 2026, Sonnet 4.8 was expected to bring significant improvements inherited from the Opus 4.7 architecture, most notably 3.75MP image resolution support, a new 'xhigh' effort level, and a substantially improved coding capability projected at 82-84% on SWE-bench Verified—performance that would have placed it ahead of Opus 4.5 (80.9%) and Opus 4.6 (80.8%) at just 60% of Opus pricing.
The version numbering is notable: Anthropic skipped version 4.7 for the Sonnet line, suggesting Sonnet 4.8 originated from a distinct training run and development timeline rather than being a simple derivative of Opus 4.7. This architectural divergence likely reflects optimizations specific to the cost-performance sweet spot that defines the Sonnet tier.
However, the landscape shifted rapidly. By late June 2026, Anthropic released Claude Sonnet 5 as the definitive successor to Sonnet 4.6, achieving 85.2% on SWE-bench Verified, 63.2% on SWE-bench Pro, and 80.4% on Terminal-Bench 2.1 at the same $3/$15 pricing (with introductory pricing of $2/$10 through August 2026). This suggests Sonnet 4.8 may have served as an internal stepping stone or was briefly available before being absorbed into the Sonnet 5 release. The most current version of the mid-tier Claude model is Claude Sonnet 5.
벤치마크 및 성능
Based on leaked specifications and extrapolation from the Opus 4.7 improvement trajectory, Sonnet 4.8 was projected to deliver substantial benchmark gains over its predecessor:
| Benchmark | Sonnet 4.6 (Actual) | Sonnet 4.8 (Expected) | Improvement |
|---|---|---|---|
| SWE-bench Verified | 79.6% | 82-84% | +3-4 points |
| GPQA Diamond | 74.1% | 76-78% | +2-4 points |
| Image Resolution | ~1.25MP | 3.75MP | 3x increase |
| Context Window | 1M (beta) | 1M (possible GA) | Stabilized |
For context, the SWE-bench Verified leaderboard as of late May 2026 showed:
| Rank | Model | Score |
|---|---|---|
| 1 | Claude Mythos Preview | 93.9% |
| 2 | Claude Opus 4.7 | 87.6% |
| 3 | GPT-5.3 Codex | 85.0% |
| 4 | Claude Opus 4.5 | 80.9% |
| 5 | Claude Opus 4.6 | 80.8% |
| 6 | DeepSeek V4 Pro(Max) | 80.6% |
| 7 | Gemini 3.1 Pro | 80.6% |
| 8 | Kimi K2.6 | 80.2% |
| 10 | Claude Sonnet 4.6 | 79.6% |
At 82-84%, Sonnet 4.8 would have surpassed both Opus 4.5 and Opus 4.6 on SWE-bench Verified, delivering Opus-level coding at Sonnet pricing. The subsequent Sonnet 5 release (85.2% SWE-bench Verified, 63.2% SWE-bench Pro) appears to have validated and exceeded these projections.
**Tokenizer impact**: A critical operational consideration is the updated tokenizer (shared with Opus 4.7), which generates 1.0x-1.35x more tokens depending on content type:
| Content Type | Token Increase |
|---|---|
| English Prose | ~1.0x (unchanged) |
| Code | ~1.1-1.2x |
| Structured Data (JSON/XML) | Up to 1.35x |
This means effective costs could rise 10-35% even at unchanged API pricing.
상세 비교
### Claude Sonnet 4.8 vs Sonnet 4.6
Sonnet 4.8 represents an incremental but meaningful upgrade over Sonnet 4.6. The expected 3-4 point gain on SWE-bench Verified (82-84% vs 79.6%) is compounded by a 3x improvement in image resolution (3.75MP vs ~1.25MP), making it substantially more capable for document analysis, UI mockup understanding, and chart interpretation. Both share the same $3/$15 pricing tier and 1M context window. The main trade-off is the new tokenizer, which could increase effective costs by 10-35% for code and structured data.
### Claude Sonnet 4.8 vs GPT-5.2
At $1.25/$10 per million tokens, GPT-5.2 is significantly cheaper than Sonnet 4.8's $3/$15. GPT-5.2 scores approximately 80.0% on SWE-bench Verified, placing it within the lower range of Sonnet 4.8's expected 82-84%. However, Sonnet 4.8's advantage lies in its superior vision capabilities (3.75MP), 1M context window, and the 'xhigh' effort level for nuanced performance tuning. For cost-sensitive workloads, GPT-5.2 offers a compelling 2.4x cheaper alternative with comparable coding performance.
### Claude Sonnet 4.8 vs DeepSeek V4 Pro(Max)
DeepSeek V4 Pro(Max) at $1.74/$3.48 is the most affordable frontier model, achieving 80.6% on SWE-bench Verified at roughly one-quarter the output cost of Sonnet 4.8. DeepSeek leads on raw cost efficiency, while Sonnet 4.8 is expected to lead on absolute coding performance by 2-3 points. Sonnet 4.8's ecosystem integration (Claude Code, MCP connectors, 1M context at standard pricing) and the effort-level control system provide additional value that DeepSeek's pricing cannot match.
### Claude Sonnet 4.8 vs Opus 4.7
Opus 4.7 at $5/$25 achieved 87.6% SWE-bench Verified, maintaining a 4-5 point lead over Sonnet 4.8's expected range. Opus 4.7 is 67% more expensive on output tokens. The gap narrows substantially on vision tasks (both share 3.75MP and 98.5% visual accuracy), but Opus retains clear advantages on deep reasoning and complex multi-file engineering. Sonnet 4.8 offers the better value proposition for teams that don't need the absolute ceiling.
### Key Specifications Comparison
| Spec | Sonnet 4.8 | Sonnet 5 | Opus 4.8 |
|---|---|---|---|
| Release | Expected May 2026 | June 30, 2026 | May 28, 2026 |
| Context | 1M | 1M | 1M |
| Max Output | ~128K | 128K | 128K |
| Input Price | $3 | $3 ($2 intro) | $5 |
| Output Price | $15 | $15 ($10 intro) | $25 |
| SWE-bench Verified | 82-84% (exp.) | 85.2% | 88.6% |
| Effort Levels | 5 (incl. xhigh) | 5 (incl. xhigh) | 5 (incl. xhigh) |
커뮤니티 평가
Claude Sonnet 4.8 occupied an unusual position in the developer community—it was heavily anticipated based on leaks but appears to have been either a brief intermediate release or an internal stepping stone toward Sonnet 5.
**Anticipated reception (pre-release, May 2026)**: Developers following the leaks on AI Models Navi and related channels expressed strong interest in the projected 82-84% SWE-bench performance at Sonnet pricing. Key developer hopes consolidated around:
- More reliable long-form output generation
- Reduced over-engineering (less unnecessary complexity added to solutions)
- Improved Time-to-First-Token (TTFT)
- Better multi-file awareness in large codebases
- Stable JSON/structured output with constrained decoding
**Post-Sonnet 5 reception (June-July 2026)**: When Sonnet 5 launched as the de facto successor, community reaction was mixed. On Hacker News and X, developers praised the benchmark improvements and introductory pricing ($2/$10), with one commenter calling it "another great incremental update to the workhorse." However, several noted that at full standard pricing ($3/$15), the value proposition was "far more compelling" at the launch price than long-term. Some skepticism emerged about whether Sonnet 5 justified the upgrade from Sonnet 4.6, with one user noting "if you're doing something hard, just use a bigger model."
**Pricing concerns**: The new tokenizer's 10-35% token inflation drew attention, with developers noting that even unchanged per-token pricing effectively means higher costs for code-heavy workloads. This was partially offset by Anthropic's promotional pricing through August 2026.
**Claude Code integration**: Sonnet 5 (the successor) became the default model for Claude Code, and rate limits were doubled to 4,000 requests/minute for Tier 4 accounts. This operational improvement was well-received by teams running high-volume agentic workflows.
활용 사례
### 1. High-Volume Agentic Coding Workflows
Sonnet 4.8 was positioned to be the optimal model for teams running hundreds of autonomous coding agents in parallel. At $3/$15 per million tokens (vs Opus's $5/$25), running 100M output tokens monthly would cost ~$1,500 vs ~$2,500 on Opus—a $1,000/month savings per workflow pipeline. With projected 82-84% SWE-bench performance, it would handle the vast majority of bug fixes, refactoring tasks, and feature implementations without needing Opus-level reasoning. **Choose over alternatives when**: you're running high-volume automated code review, CI/CD integration testing, or bulk code migration tasks where the 3-4 point gap to Opus doesn't justify 67% higher output costs.
### 2. Document and Visual Analysis at Scale
The jump from ~1.25MP to 3.75MP image resolution makes Sonnet 4.8 significantly more practical for enterprise document processing. Use cases include analyzing high-resolution PDF contracts, interpreting architectural diagrams and UI mockups, extracting data from complex financial charts, and processing scanned forms. At 3x the resolution of Sonnet 4.6 with 98.5% visual accuracy (inherited from Opus 4.7), it can handle documents that previously required Opus-tier models. **Choose over alternatives when**: processing high-resolution visual documents is a core requirement and you need Opus-quality vision without Opus pricing.
### 3. Multi-Step Tool Orchestration with Fine-Grained Control
The new 'xhigh' effort level (between 'high' and 'max') gives developers unprecedented control over the cost-performance tradeoff in tool-use scenarios. For agent pipelines that coordinate multiple MCP servers, manage complex state transitions, or execute multi-step workflows, xhigh provides a middle ground—more thorough reasoning than 'high' without the full cost of 'max.' Combined with the expected improvements in tool use, this makes Sonnet 4.8 suitable for customer support automation, data pipeline orchestration, and CRM integration agents. **Choose over alternatives when**: you need to tune effort levels granularly and want a single model that scales from quick lookups (low effort) to complex multi-step planning (xhigh).
### 4. Cost-Optimized Production with Prompt Caching
At the Sonnet price tier, combining the 1M context window with Anthropic's prompt caching (90% discount on cache reads) makes Sonnet 4.8 highly cost-effective for applications that repeatedly process similar context—legal document analysis, codebase-wide refactoring, and knowledge base querying. With the beta 1M context potentially moving to GA, teams can load entire repositories or document collections in single requests, reducing per-query overhead. **Choose over alternatives when**: your workload involves repeated analysis of large, semi-static contexts where caching can offset the 10-35% tokenizer inflation.
최신 뉴스
**Sonnet 4.8 Status (as of July 2026)**: Claude Sonnet 4.8 appears to have been either a brief intermediate release or an internal development milestone. The Sonnet line advanced directly from 4.6 to Sonnet 5, which launched June 30, 2026 as the current mid-tier model.
**Claude Sonnet 5 (Successor) - Released June 30, 2026**:
- SWE-bench Verified: 85.2%, SWE-bench Pro: 63.2%, Terminal-Bench 2.1: 80.4%
- Introductory pricing: $2/$10 through August 31, 2026 (standard: $3/$15)
- 1M token context window, 128K max output
- 5 effort levels including 'xhigh'
- 23% faster inference than Sonnet 4.7
- Default model for Free and Pro Claude plans
- Rate limits doubled to 4,000 req/min for Tier 4 accounts
**Claude Opus 4.8 - Released May 28, 2026**:
- SWE-bench Verified: 88.6%, USAMO 2026: 96.7%
- Pricing: $5/$25 (fast mode: $10/$50 at 2.5x speed)
- 1M context, 128K max output
- New 'xhigh' effort level added
**Deprecation timeline**: Anthropic deprecated Claude 3.5 and 4.6 model families as of June 30, 2026. Claude 4.7 Sonnet remains supported until December 2026.
**Anthropic promotional pricing**: Sonnet 5 available at $2/$10 (intro) through August 31, 2026, making it temporarily 60% cheaper than Opus 4.8.